IEEE/ACM TRANSACTIONS ON COMPUTATIONAL BIOLOGY AND BIOINFORMATICS, VOL. XX, NO.

XX, XXXX 2023 1

FelAug and WaterAug: Superpixel-Based Medical Image Augmentation

Methods for Reducing False Predictions In MRI Image Classifications
Subhendu Sinha Chaudhuri, Saniran Basak, Bhaskar Mazumder, and Sheli Sinha Chaudhuri Member,
IEEE

Abstract— Although deep learning is gaining importance in approval from Human Ethics Committees, limited patient popula-
medical image classification, the nonavailability of diverse training tions for certain conditions, the reluctance of patients to consent
data is the greatest challenge in developing robust networks. A
to image collection, and difficulties in accessing patients across
significant challenge in automated disease detection is the high
rate of false predictions, both false positives and false negatives, diverse geographic locations. Unlike natural images, labelling medical
which threaten patient’s conditions. False positives expose the images also requires assistance from medical experts, which is
patients to unnecessary medication, while false negatives deprive time-consuming. These problems give rise to less data, resulting in
the patient of any medication. In this work, we developed two new overfitting during network training. However, medical data also suffer
augmentation techniques, FelAug and WaterAug, based on the
Felzenszwalb and Watershed superpixel segmentation methods. from class imbalance. Data imbalance results in biases in model
The images were segmented using the two methods, and then training, leading to inaccurate predictions [3].
individual segments were changed to produce a new augmented Researchers often employ data augmentation techniques to over-
image. Each image produced two augmented images. We trained come the scarcity of real-life data [4]. One or a combination of
four deep-learning network networks with images produced by our
model. Our method significantly reduced false positive and false
methods, such as random rotation, flipping, cropping, zooming, and
negative cases compared to the network trained with classically others, achieves classical image augmentation. Keras [5] uses a
augmented images. We also justified the results by showing that function called ImageDataGenerator to implement data augmentation,
FelAug and WaterAug methods produce more feature maps in performing it on the fly.
augmented images.
Superpixels refer to a collection of clustered pixels due to their
Index Terms— Deep Learning, Superpixel Augmentation, shared characteristics, such as colour, spatial proximity, and other
Medical Imaging, False Predictions properties. X. Ren and J. Malik first introduced superpixels [6]
in their “Learning a Classification Model for Segmentation” paper.
Consequently, researchers have developed several other methods
I. I NTRODUCTION
for superpixel segmentation. Superpixels enhance the computational
Radiologists and clinicians analyze various medical images to efficiency of subsequent image-processing tasks by reducing data re-
diagnose a particular disease easily. The complexity of the disease dundancy while preserving image structural information. Researchers
and variations in the opinions of experts sometimes make accurate have employed superpixel segmentation as a preprocessing technique
diagnosis difficult. Semi-automatic and automated systems assisted for object segmentation [7] [8] [9] anomaly detection [10], semantic
by AI help experts make fast and accurate decisions. Deep learning segmentation [11], image classification [12] [13] [14], and saliency
(DL) –based models for medical imaging are gaining importance [1] detection [15] [16]. Superpixels have also been used with Deep
[2]. All DL and CNN (convolutional neural network)-based models Learning and Convolutional Neural Networks to model different
have two associated prediction factors: False Positive and False image preprocessing tasks, such as image segmentation [17] [18].
Negative. A disease can be falsely predicted as a false positive,
meaning the patient can receive treatment even if they are not
suffering from the same disease. A patient can also be diagnosed II. L ITERATURE S URVEY
with false negatives, meaning that although he is suffering from a A comprehensive survey of the superpixelation technique, its
disease, he is diagnosed with no disease condition and thus receives challenges, applications, and commonly used evaluation metrics in
no medical attention. Both of these situations can be life-threatening. the paper ’A survey of superpixel methods and their application’
All automated diagnosis systems should work to achieve zero false [19]. The review work on clustering techniques for the superpixel
positive and false negative results. method was performed by Zhijie Chen et al. [20], who compared the
To achieve high accuracy, researchers should train artificial intelli- advantages and disadvantages of different methods. Our literature
gence systems with a large amount of diverse data to make them survey includes some papers related to our work.
robust and versatile. A significant challenge in utilizing medical Achanta used the K-means clustering algorithm in 2010 to propose
images and data, however, is the issue of data scarcity. This scarcity SLIC (Simple Linear Iterative Clustering) [21], a method for gener-
stems from various factors, including the necessity of obtaining ating superpixels. Various modifications to SLIC have been made,
including Linear Spectral Clustering (LSC) [22], FLIC (Fastlinear
This paragraph of the first footnote will contain the date on which you
submitted your paper for review. iterative clustering) [23], and SNIC (Simple noniterative clustering)
Subhendu Sinha Chaudhuri is with the Department of Pharmaceuti- [24]. Zheng et al. developed Fuzzy SLIC (Fuzzy Simple Linear
cal Sciences, Dibrugarh University, Dibrugarh-786004, Assam, India(e- Iterative Clustering) [25].
mail: [email protected]). Graph-based superpixel algorithms treat every image as an undi-
Samiran Bask is with the Department of Electronics and Telecom-
munication Engineering, Jadavpur University, Jadavpur -700032, West rected graph, dividing the graph into segments based on the features
Bengal, India (e-mail: [email protected]). of its edges. Normalized Cut (NCut) [26]developed by Ren and
Bhaskar Mazumder is with the Department of Pharmaceutical Sci- Mallick uses this method and developed a superpixel algorithm.
ences, Dibrugarh University, Dibrugarh-786004, Assam, India(e-mail: Merging pixels from bottom to top in an undirected graph results
[email protected]).
Sheli Sinha Chaudhuri is with the Department of Electronics
in an ERS (Entropy Rate Superpixel) [27]. Vargas-Munoz et al
and Telecommunication Engineering, Jadavpur University, Jadavpur - devised an algorithm for creating superpixels by initiating seed
700032, West Bengal, India (e-mail: [email protected]. growth in a sequential series of image forecasting transforms [28]. In
2 IEEE/ACM TRANSACTIONS ON COMPUTATIONAL BIOLOGY AND BIOINFORMATICS, VOL. XX, NO. XX, XXXX 2023

their paper ’Efficient graph-based image segmentation’, Felzenszwalb it suitable for medical image segmentation. Watershed superpixel
and Huttenlocher [29]developed a superpixel segmentation algorithm segmentation is a robust image segmentation method independent
based on pairwise region comparison. Their method preserved details of image content. The watershed segmentation method also tends to
in low-variability image regions while ignoring details in high- preserve the structural information of the image [31], which is vital
variability image regions. in the case of medical images.
In 1978 Digabel and Lantuejoul [30] introduced a region-based We modified both segmentation algorithms to produce new aug-
image segmentation method known as Watershed segmentation. This mented images to increase the size of the training image dataset.
technique was applied to segment CT images of lymph nodes [32] In the FelAug method, we randomly chose 10% 20% and 30% of
and yielded minimum errors in phantom area measurements and inter- segments done with Felzenszwalb’s method and replaced each pixel
operator variations among five different segmentation techniques The of those segments with the mean pixel value of that segment. Hence,
spatial-constrained Watershed (SCoW) superpixel algorithm [33] was we obtained three image sets from a single image.
developed based on the watershed algorithm. ScoW works by placing We also applied a similar technique to obtain images using the
evenly placed controlled markers. Other superpixel algorithms devel- WaterAug method, Watershed’s segmentation method, instead of
oped based on watershed segmentation methods include Compact Felzenszwalb’s.
Watershed, Morphological Superpixel Segmentation, Waterpixel, and After augmentation, each new training set contained 11,424 im-
Deep Watershed. ages, 5712 images derived by the FelAug method and the remaining
We have also studied different augmentation methods used to images derived by the WaterAug method. We have shown the
augment MRI images only, as we have done this work with MRI structure of the Augmented training dataset in Fig 1
images of brain tumours. Khan et al. [34] added noise and sharpened The algorithms of Felzenszwalb’s method is given in Section IV-A.
the MRI images for image augmentation. In a benchmarking study on Section IV-B compares the image characteristics between the original
3D CNNs, Dufumier B et al. [35] also evaluated the benefits of data image and the augmented images
augmentation and deep ensemble learning. He used data augmen-
tation techniques, including translation, blurring, random cropping, A. Augmentation Algorithm
rotation, and noise injection. Other researchers [36] and [37] applied The FelAug algorithm is given in Algorithm 1. The WaterAug
random scaling, rotation, and elastic deformation techniques to gen- method uses a similar augmentation technique, applying the Water-
erate augmented images. Researchers have also employed Generative shed segmentation method.
Adversarial Networks (GANs) to create synthetic MRI images [38]. Figure 2 shows segments chosen randomly with FelAug and the
Diana et al. [39] showed that in the case of most of the augmented corresponding augmented images. Fig2a is the original image. Fig 2b
methods, reducing the FP count resulted in a reduction in the TP. shows the segments formed with Felzenszwalb segmentation. The red
Zhang et al. [40] developed superpixel-based data augmentation areas in Figures 2c, 2d, 2e, and 2f represent the randomly selected
(SPDA) for medical images. They used the SLIC [21] superpixel 10%, 20%, 30%, and 40% of the Felzenszwalb-segmented regions,
algorithm to generate superpixels. SuperpixelGridCut, SuperPixel- respectively. The algorithm modifies the pixel values in these selected
GridMean, and SuperpixelGridMix [41] were used to develop Su- regions to µI .The algorithm generates Figures 2g, 2h, 2i, and 2j by
perpixelGridMasks data augmentation. replacing every I(i, j) with µI . The operation resulted in 10%, 20%,
D. Sun and F. Dornaika [42] used a superpixelwise image fusion 30% and 40% FelAug augmented images.
approach for image augmentation. Researchers applied this method to Figure 3a shows the original image segmented with the watershed
nonmedical images, producing augmented images aware of the local segmentation technique.In Figures 3b,3c,3d, and 3e, we represent the
context. This method produced better image classification results and red areas as randomly selected portions comprising 10%, 20%, 30%,
outperformed many region-based augmentation methods. and 40% of the Watershed segmented area, respectively, and modify
their pixel values to µI . In Figures 3f, 3g, 3h, and 3i, we form the
III. E QUIPMENTS images by replacing every I(i,j) with I(i, j) with µI . The operation
The MRI dataset was sourced from a publicly accessible collection resulted in 10%, 20%, 30% and 40% WaterAug augmented images.
of brain tumour images, accessible at the following https:
//www.kaggle.com/datasets/masoudnickparvar/ B. Discussion on Augmented Images
brain-tumor-mri-dataset/data. We compared the augmented images with the original image using
The computer was a Ryzen 5 3600, with 32 GB of RAM and a standard procedures. Functions for Mean Square Error (MSE) and
GTX 3060 12 GB graphics card. Peak Signal-to-Noise Ratio (PSNR) actively compared two images
Operating System: WSL (Windows Subsystem for Linux) at the pixel level. We used the Structural Similarity Index(SSIM)
Software: TensorFlow with other modules. and Normalized Cross Relation to compare the structural differences
between the two images.
IV. M ETHODOLOGY We also compared the images using statistical metrics tools such as
We used an MRI dataset with 7023 images of human brain Entropy Difference and Mutual Information functions. We also com-
tumours classified into four types: glioma, meningioma, no tumour, pared Augmented images with the original image using two feature-
and pituitary. The dataset consists of 5712 images in the training based comparison methods: Scale-Invariant Feature Transform(SIFT)
folder and 1311 in the testing folder, divided into four subclasses. and histogram comparison. Table I compares the original image and
Contrast Limited Adaptive Histogram Equalization(CLAHE) was images obtained with the FelAug technique. Table II compares the
applied to all the training images as a preprocessing step to increase original image and images obtained with the WaterAug technique.
The mathematical expression for calculating MSE is
the contrast of the images without increasing the noise. Each image
was then segmented using the Graph-based Superpixel Segmentation N
1 X
method proposed by Felzenszwalb [29]and Watershed Segmentation MSE = (I1 (i) − I2 (i))2
N
[30]. Felzenszwalb’s method preserves image details in low vari- i=1
ability regions while remaining computationally efficient, making where
AUTHOR et al.: PREPARATION OF BRIEF PAPERS FOR IEEE TRANSACTIONS AND JOURNALS (DECEMBER 2023) 3

Randomly Selected Augmentation

10% segments
Randomly Selected Augmentation
20% segments

Randomly Selected 10% Directory

Felzenswalb Segmentation Augmentation
30% segments

Randomly Selected Augmentation

40% segments 20% Directory
Training Dataset

Randomly Selected Augmentation 30% Directory

10% segments

Randomly Selected Augmentation

Watershed Segmentation 20% segments 40% Directory
Randomly Selected Augmentation
30% segments

Randomly Selected Augmentation

40% segments
Fig. 1: Augmented Dataset Structure

image against an original image. The standard formula for calculating

the PSNR is

MAX2
 
PSNR = 10 · log10
MSE
(a) (b) (c) (d) (e)
where
MAX is the maximum possible pixel value of the image (e.g., 255
for 8-bit images).
MSE is the Mean Squared Error between the two images.
SSIM is calculated by the mathematical formula
(f) (g) (h) (i) (j)
Fig. 2: FelAug Augmentation (2µx µy + C1 )(2σxy + C2 )
SSIM(x, y) =
(µ2x + µ2y + C1 )(σx2 + σy2 + C2 )
where
• µx : Mean of image x.
• µy : Mean of image y.
• σx2 : Variance of image x.
• σy2 : Variance of image y.
(a) (b) (c) (d) (e)
• σxy : Covariance between x and y.
• C1 and C2 : Constants to stabilize the division, typically defined
as C1 = (K1 L)2 and C2 = (K2 L)2 , where L is the dynamic
range of pixel values (e.g., 255 for 8-bit images), and K1 and
K2 are small constants (e.g., K1 = 0.01, K2 = 0.03).
(f) (g) (h) (i) PSNR and SSIM values also show minimum structural similarity
Fig. 3: WaterAug Augmentation with 40% pixel augmentation and maximum similarity with 10%
pixels. Data from Table I and Table II shows that there is a small
difference in PSNR and SSIM values between the FelAug and the
WaterAug images. This shows that although the images are not very
• N : Total number of pixels in the images.
structurally different from the original image, they are also not the
• I1 (i): The pixel value at the i-th position in the first image.
same.
• I2 (i): The pixel value at the i-th position in the second image.
Normalized Cross Relation (NCR) values between two images
Higher MSE values indicate more significant dissimilarity between indicate the degree of similarity in their pixel intensity distributions,
the original and augmented images. Tables I and II show that the normalized to account for variations in scale. It quantifies how well
10% column has the lowest MSE values, and the 40% column has the pixel intensities in one image correlate with those in another,
the highest value, indicating that the 10% image is almost similar to making it a useful metric for assessing the alignment or consistency
the original image and the 40% image has the highest dissimilarity. between two images. An NCR near one indicates high similarity,
The PSNR value is used to compare a reconstructed or processed whereas zero or negative values indicate no similarity. NCR is
4 IEEE/ACM TRANSACTIONS ON COMPUTATIONAL BIOLOGY AND BIOINFORMATICS, VOL. XX, NO. XX, XXXX 2023

Algorithm 1 FelAug scans. Unlike pixel-by-pixel similarity metrics like MSE or SSIM, MI
1: function A LGORITHM F OR AUGMENTATION(input image) emphasizes the statistical relationships between the intensity values.
2: Read the input image Mutual information is calculated by
3: Apply Contrast Limited Adaptive Histogram Equalization (CLAHE) to enhance
contrast
4: Segment the image using the Felzenszwalb method  
5: procedure
XX pI1 ,I2 (i, j)
6:
MI(I1 , I2 ) = pI1 ,I2 (i, j) log
Initialize an empty priority queue Q pI1 (i) · pI2 (j)
7: Initialize an empty disjoint-set data structure i j
8: Sort all edges in non-decreasing order of their weights
9: for each edge (u, v) in sorted order do where
10: Find the root of the set containing u (root u) and the root of the set
containing v (root v) • pI1 ,I2 (i, j): The joint probability distribution of pixel intensities
11: if root u ̸= root v then i in image I1 and j in image I2 .
12: Merge the sets containing u and v
• pI1 (i): The marginal probability distribution of pixel intensities
13: Add the edge weight to Q
14: end if in image I1 .
15: end for • pI2 (j): The marginal probability distribution of pixel intensities
16: Find the maximum edge weight in Q, denoted as max weight
17: Initialize a threshold parameter, scale, as max weight divided by the in image I2 .
number of vertices in the image From both Tables I and II, we can see that 10% augmented
18: Initialize an empty list to store segments
19: for each edge (u, v) in sorted order do images by both FelAug and WaterAug method have a value of
20: Find the root of the set containing u (root u) and the root of the set around 4.1, suggesting a strong relationship between the original
containing v (root v)
21: if root u ̸= root v and the edge weight ≤ scale × max weight
and augmented images. Both methods 20% and 30%, augmentation
then techniques have an MI value of around 3. With 40% superpixel
22: Merge the sets containing u and v augmentation, the value drops around 2.5, showing an evident decline
23: else
24: Add the edge weight to Q in similarity between the original and augmented images as the
25: end if augmentation increases. Augmentation by Rotation shows a drastic
26: end for
decrease to 0.2860, indicating that Rotation severely reduces the
27: for each pixel in the image do
28: Assign its label to the root of the set containing it mutual information between the original and augmented images.
29: end for Entropy Difference with the Original Image quantifies an image’s
30: end procedure
31: Randomly select 20% of the segments randomness or information content change after augmentation or
32: for each selected segment do modification. It compares the entropy (a measure of uncertainty or
33: Initialize a variable to store the sum of pixel values and a variable to count complexity) of the original image with the entropy of the modified
the number of pixels
34: for each pixel in the segment do or augmented image. Entropy is represented by
35: Add the pixel value to the sum
36: Increment the pixel count X
37: end for H=− P (i) log2 P (i)
38: Calculate the mean value of the pixels in the segment by dividing the sum i
by the count
39: for each pixel in the segment do Where
40: Replace the pixel value with the calculated mean value
41: end for P (i): Probability of the ith intensity value in the image.
42: end for H: Entropy of the image.
43: Display or save the processed image
A slight entropy difference indicates minimal alteration of the
44: end function
pixel intensity distribution. Tables I and II show that 10% pixel
augmented image has the minimum entropy difference while 40%
pixel augmentation has the highest.
calculated using the formula
P The histogram of an image represents the distribution of the
i,j I1 (i, j) · I2 (i, j) pixel intensities. The histogram correlation indicates the degree of
NCR = r  P  correlation between the histograms. Since we have applied CLAHE
2 (i, j) · 2 (i, j)
P
I
i,j 1 I
i,j 2 to the images before applying FelAug or WaterAug, the values near
zero show that the histograms of the two images are not the same.
Where As no histogram equalization was applied in the case of the rotated
I1 (i, j): The pixel intensity value of the first image at position image, the histogram correlation value is near 1
(i, j). SSIM, Mutual Information, and Histogram Correlation values of
I2 (i, j): The pixel intensity value of the second image at position both tables show that images formed with 10% of the augmented
(i, j). P  segments are more similar to the original image, and 40% of selected
The summations i,j run overall image pixel coordinates. augmentation are the most dissimilar.
From Table I, we can see that the maximum performance is We used the SIFT (Scale-Invariant Feature Transform) algorithm
obtained with 10% augmented pixels, and as the percentage of to detect and describe images’ local features (key points). These key
augmentation increases, the performance of the technique decreases points are highly distinctive and invariant to scaling, rotation, and
slightly. The score for rotation-based augmentation is 0.2360, which illumination changes, making SIFT a robust tool for applications like
is significantly lower than pixel-based augmentation. The NCR values image matching. From Table Iand Figure 4, we see that the original
from both Tables I and II shows that, with superpixel based segmenta- image had 2340 key points, and the image formed with 20% selected
tion technique, the images retain their similarity to the original image. segments augmented with the FelAug method has 2246 key points
Since 40% of augmented pixels have a comparatively low NCR of with 1692 matched key points. This means that the FelAug method
0.98, it is less similar to the original image. generated 554 new key points.
Mutual Information is especially useful for comparing images Similarly, data from Table IIand Figure 5 shows that when aug-
with varying intensity scales, such as multimodal images like MRI mented by WaterAug methods, the same image generated 2159
AUTHOR et al.: PREPARATION OF BRIEF PAPERS FOR IEEE TRANSACTIONS AND JOURNALS (DECEMBER 2023) 5

Serial Name of the Method 10% Pixels Aug- 20% Pixels Aug- 30% Pixels Aug- 40% Pixels Aug- Augmentation by
No mented mented mented mented Rotation
1 MSE 5.4396 11.6714 15.3122 23.1910 104.2130
2 PSNR 40.7752 37.4596 36.2804 34.4776 27.9516
3 SSIM 0.9761 0.9508 0.9346 0.8950 0.1230
4 Normalized Cross Rela- 0.9968 0.9921 0.9904 0.9850 0.2360
tion
5 Mutual Information 4.1092 3.4093 3.1006 2.5880 0.2860
6 Entropy Difference with 0.0099 0.0165 0.0252 0.0566 0.5594
Original Image
7 Histogram Correlation 0.9903 0.9268 0.9603 0.9411 0.1442
5 SIFT (Fig 4)
No. of key points in the 2340 2340 2340 2340 2340
original image (Image 1)
No of key points in the 2348 2246 2134 2033 1891
matched image (Image 2)
Matched key points 1959 1692 1488 1272 1464
New Points Generated 389 554 646 761 427

TABLE I: Performance Metrics for FelAug Augmentation Techniques

Serial Name of the Method 10% Pixels Aug- 20% Pixels Aug- 30% Pixels Aug- 40% Pixels Aug-
No mented mented mented mented
1 MSE 5.4148 12.6024 19.6322 25.0987
2 PSNR 40.7949 37.1263 35.2011 34.1343
3 SSIM 0.9690 0.9294 0.8874 0.8554
4 Normalized Cross Corre- 0.9946 0.9854 0.9797 0.9734
lation
5 Mutual Information 4.1569 3.4283 2.8597 2.1599
6 Entropy Difference with 0.0027 0.0153 0.0212 0.0284
the Original Image
7 Histogram Correlation 0.9970 0.9912 0.9871 0.9750
5 SIFT (Fig 5)
No. of key points in the 2340 2340 2340 2340
original image (Image 1)
No of key points in the 2251 2159 2017 1856
matched image (Image 2)
Matched key points 1951 1551 1268 1029
New Points Generated 300 608 749 827

TABLE II: Performance Metrics for WaterAug Augmentation Techniques

images with 1551 matched key points. This means that the WaterAug
method generated 608 new key points.
The data in column 7 of both tables shows that with rotated images,
the number of key points is less, with more matched key points,
indicating less key point generation during augmentation by rotation.
Tables Iand II show that augmented images formed using 10% seg-
ment selection generate very few new key points. This and previously
discussed points made 10% of augmented images unsuitable.
Column 6 of Tables Iand II demonstrates that 40% of augmented
images generate many new key points. However, since 40% aug-
mented images showed a high MSE and entropy difference with a
low SSIM compared to the original image, we did not consider 40%
Fig. 4: SIFT Comparison FelAug 20% augmented for training the model.

C. Database Development
The algorithm created two training dataset directories of augmented
images, each containing two augmented images for each original
training image—one created using the FelAug method and the other
using the WaterAug method. One directory contained images with
20% augmented pixels and the other with 30% augmented pixels.
Fig.6 shows the directory structure. Another training dataset included
the original images, which the system augments on the fly during
training using classical augmentation techniques of random flipping,
Rotation by 40 degrees, zoom, and brightness and applied separately
to modified deep neural networks. Thus, we used three different
Fig. 5: SIFT Comparison WaterAug 20%
training datasets to compare the performance of the best dataset.
6 IEEE/ACM TRANSACTIONS ON COMPUTATIONAL BIOLOGY AND BIOINFORMATICS, VOL. XX, NO. XX, XXXX 2023

Fig. 6: Training and Testing Dataset Structure

We used MobileNetV2, InceptionV3, DenseNet169 and ResNet50 classically augmented datasets, the values obtained with 20% datasets
for multiclass image classification training and validation. show slightly better results.
Each model’s final description is MobileNetV2: Total params: Cohen’s Kappa is a statistical measure used to evaluate the agree-
2,263,108, Trainable params: 2,228,996, Non-trainable params: ment between two classifiers in medical image classification tasks. It
34,112. InceptionV3: Total params: 21,810,980, Trainable params: accounts for the possibility that some agreement might occur purely
8196, Non-trainable params: 21,802,784. DenseNet169: Total pa- by chance, making it a more reliable metric than simple accuracy.
rameters 14,351,940, Trainable Parameter 14,193,540 and non- Cohen’s Kappa provides a more nuanced understanding of a model’s
Trainable:158,400. Resnet50: Total params: 23,595,908 Trainable performance by accounting for agreement beyond chance, making
params: 23,542,788, Non-trainable params: 53,120 These networks it critical in high-stakes fields like medical diagnostics. The value
were chosen based on their performance with medical image clas- of Cohen’s Kappa ranges from -1 to +1, where -1 is the complete
sification.MobileNetV2 has been used in Brain Tumor Classification disagreement (the classifier’s predictions are systematically opposite
[43], Lung Tumor Classification [44], and Skin Cancer Diagnosis to the ground truth), +1 is the complete agreement, and 0 is the
[45]. InceptionV3 has been successfully used to detect glaucoma agreement that is no better than chance. The formula to calculate
[46] and skin cancer [47]. DenseNet has been used to classify Skin Cohen’s Kappa κ
cancer [48] and Breast tumours [49]. Researchers have widely used Po − Pe
κ=
ResNet50 in medical image classification [50]. Since we aimed to 1 − Pe
compare the performance of our augmented images with classically
augmented images, we did not perform any hyperparameter tuning Where:
of the networks • Po : The observed agreement, representing the proportion of
We tested each transfer learning model on three datasets: Training times the model’s predictions agree with the ground truth.
dataset 1 contained images formed with 30% of augmented segments • Pe : The expected agreement, representing the proportion of
randomly selected. Dataset 2 contained images with 20% of the agreement expected to occur by random chance (calculated
randomly selected augmented segments. Dataset 3 consisted of the based on the distribution of predicted and actual labels).
original training dataset, which will be augmented on the fly and DenseNet169 achieves a kappa value of κ = 0.978 for size 30 and
labelled Non-Aug in Section V. κ = 0.980 for size 20, demonstrating near-perfect agreement. It even
delivers the best kappa values with a non-augmented training dataset,
which the system augments on the fly. ResNet50 also shows a high
V. R ESULTS AND D ISCUSSION
kappa value κ = 0.961 for 30 and κ = 0.963, indicating excellent
Tables III,IV, and V present the training and testing results for all agreement with the ground truth. MobileNetV2 and InceptionV3
three datasets on four CNN models. perform moderately.
Table III shows that Validation and Test Accuracy are low for Table IV shows a drastic reduction of False Positive and False
training datasets with augmentation on the fly for all four transfer Negative predictions in all disease classes when using the FelAug and
learning models. The values also show that DenseNet169 is the WaterAug image datasets as training datasets. DenseNet169 shows
best transfer learning model for multiclass MRI image classification. the best performance with both 20% and 30% datasets. ResNet50
Although 20% and 30% superpixel augmentation datasets outperform also shows a sound reduction in FP and FN.
AUTHOR et al.: PREPARATION OF BRIEF PAPERS FOR IEEE TRANSACTIONS AND JOURNALS (DECEMBER 2023) 7

Since we have an unbalanced training dataset, we use MCC

(Matthews Correlation Coefficient) to evaluate the quality of mul-
ticlass classifications. MCC considers true positives (TP), true neg-
atives (TN), false positives (FP), and false negatives (FN), making
it a balanced measure even when class distributions are imbalanced.
MCC is calculated using the formula

(T P · T N ) − (F P · F N )
MCC = p (a) (b)
(T P + F P )(T P + F N )(T N + F P )(T N + F N )
The value of MCC range from +1 to -1 , where +1 indicates perfect
prediction and -1 indicating completely wrong prediction.
GMean also evaluates the balance between sensitivity and speci-
ficity across all classes. This value is also particularly useful for
imbalanced datasets. The Gmean value extends from 1 to 0, where 1
indicates perfect classification, and zero indicates worst classification.
GMean is evaluated using the formula
v
u n (c) (d)
uY
n
G-Mean = t Sensitivityi
i=1

Hamming loss calculates the fraction of incorrectly predicted

labels. Haming loss is calculated using
N L
1 XX
Hamming Loss = ⊮(yij ̸= ŷij )
N ·L
i=1 j=1

N is the total number of samples. L is the total number of labels. (e) (f)
⊮ is the indicator function, which is 1 if yij ̸= ŷij , and 0 otherwise. Fig. 7: AUC ROC Curve
Table V shows that DenseNet169 achieves the highest values for
both GMean and MCC across all three training datasets. Training
with 20% and 30% superpixel augmentation datasets performs better AUC values vary between 0.69 and 0.96. The ROC curves show
than the training dataset formed by augmenting images on the fly. The more spread, meaning there is more False Positive and False Negative
20% augmented dataset with a GMean value of 0.989 and MCC value detection. The ROC curve of Fig 7a and Fig 7d curve suggests that
of 0.979 shows slightly better performance than the 30% augmented both models’ performance is suboptimal, as the curves are not as
dataset. close to the top-left corner as in the augmented cases. The ROC
ResNet50 also performs well in MRI image classification, as seen curve is less steep and does not reach the top-left corner, indicating
from the different MCC, GMean and Hamming Loss values in Table lower true positive rates (TPR) and higher false positive rates (FPR).
V. The model struggles to distinguish between classes with on-the-fly
Since False positives and false negatives play an essential role augmentation.
in medical image diagnosis, the F1 score is a crucial metric for Fig 7c and Fig 7f are the ROc-AUC curves of the 20% augmented
evaluating model performance. The F1 score, as a harmonic mean dataset obtained with DenseNet169 and ResNet50, respectively. The
of precision and recall, is essential in determining performance with AUC values in both figures are nearly 1.0, showing perfect classi-
an unbalanced dataset. fication, and the curves are tightly packed near the top-left corner,
DenseNet169 achieves the highest F1 score with all 3 augmentation indicating minimal false positives. Fig 7b and Fig7e are the curves
methods, emphasizing its reliability in balancing precision and recall. obtained with a 30% augmented image dataset. The 30% image
20% augmented dataset showed the best F1 score of 0.984 with the dataset showed nearly similar results to the 20% augmented dataset.
DenseNet169 model. Both 20% and 30% of the augmented dataset
perform better than the augmentation on the fly with F1 scores.
Another vital tool for evaluating the performance of an image
VI. C ONCLUSION AND F UTURE W ORK
classification model is the Receiver Operating Characteristic (ROC) Table III, shows that DenseNet169 performs the best among all four
curve and Area Under the Curve (AUC). ROC curve plots the transfer learning models. Densenet169 showed the best validation
TPR(True Positive rate) against the False positive rate at various and test accuracy results with all three training datasets. Resnet50
threshold levels. AUC represents the area under the ROC curve, performs closely to DenseNet169. Table IV, and V also shows the
quantifying the model’s ability to distinguish between classes. AUC excellent performance of DenseNet169.
values range from 0 to 1, where 1 is the perfect classification,0.5 is Table III, IV, and V clearly show a reduction in FP (False Positive)
random guessing, and less than 0.5 indicates worse than random. and FN (False Negative) detections when images produced by FelAug
ROC curves help determine the optimal thresholds for decision and WaterAug are used for training. This reduction in FP and FN
boundaries, which are crucial for tasks like medical image classifica- has been arrived at without compromising test accuracy results.
tion, where the cost of false positives and false negatives may vary Results in Table IV show a significant increase in test accuracy
significantly. Since AUC is less sensitive to class imbalance than when using the FelAug and WaterAug augmented image datasets
accuracy, it is a good evaluation metric for imbalanced classes. to train the model. From Table I and II, we can see high SSIM
Fig 7a and Fig 7d is the ROC- AUC curve obtained with values for images produced with the FelAug and WaterAug models
DenseNet169 and Resnet50 for classically augmented datasets. The compared to images produced by rotation, although they have very
8 IEEE/ACM TRANSACTIONS ON COMPUTATIONAL BIOLOGY AND BIOINFORMATICS, VOL. XX, NO. XX, XXXX 2023

Model Augmentation Train Accu- Validation Test Accu- Cohen’s

Name racy Accuracy racy Kappa
ResNet50 30 0.988 0.977 0.971 0.961
20 0.993 0.987 0.973 0.963
Classical 0.508 0.496 0.477 0.263
Aug
MobileNetV2 30 0.895 0.872 0.887 0.848
20 0.995 0.988 0.978 0.970
Classical 0.855 0.839 0.841 0.785
Aug
InceptionV3 30 0.841 0.835 0.780 0.706
20 0.848 0.838 0.800 0.732
Classical 0.288 0.315 0.308 0.117
Aug
DenseNet169 30 0.995 0.983 0.983 0.978
20 0.995 0.987 0.985 0.980
Classical 0.553 0.539 0.561 0.415
Aug

TABLE III: Performance Metrics of Models with and without Segmentation

Model name Augmentation Class 0 Class 1 Class 2 Class 3

TP FP TN FN TP FP TN FN TP FP TN FN TP FP TN FN
ResNET50 30 297 23 988 3 276 4 1001 36 405 3 903 0 295 8 1003 5
20 290 11 1000 10 287 14 991 19 403 4 902 2 295 7 1004 5
Classical Aug 56 0 1011 244 1 10 999 296 402 595 311 3 157 85 926 143
MobileNetV2 30 267 59 952 33 225 31 974 81 396 36 870 9 275 22 989 25
20 276 0 1011 24 303 26 979 3 404 3 903 1 299 0 1011 1
Classical Aug 237 42 969 63 249 63 942 57 397 94 812 8 219 10 1001 81
InceptionV3 30 228 79 932 72 179 97 908 127 354 43 863 51 262 69 942 38
20 243 87 924 57 177 70 935 129 361 39 867 44 268 66 945 32
Classical Aug 42 148 146 15 22 26 210 93 21 39 207 84 23 30 247 51
DenseNet169 30 292 4 1007 8 298 14 991 8 404 1 905 1 95 3 1008 5
20 292 5 1006 8 298 11 994 8 404 0 906 1 297 4 1007 3
Classical Aug 242 255 756 58 179 280 725 127 291 33 873 114 24 7 1004 276

TABLE IV: Confusion Matrix of Models with and without Segmentation

Model name Augmentation Average of all 4 Disease Class

F1 Recall Precision Specificity MCC Gmean Hamming Loss
ResNET50 30 0.969 0.969 0.970 0.991 0.960 0.979 0.014
20 0.971 0.971 0.971 0.991 0.962 0.981 0.014
Classical Aug 0.382 0.434 0.669 0.813 0.332 0.472 0.262
MobileNetV2 30 0.880 0.880 0.885 0.962 0.845 0.919 0.056
20 0.977 0.976 0.978 0.993 0.970 0.984 0.011
Classical Aug 0.835 0.828 0.853 0.946 0.787 0.883 0.079
InceptionV3 30 0.770 0.773 0.769 0.927 0.698 0.845 0.110
20 0.788 0.793 0.789 0.934 0.725 0.858 0.100
Classical Aug 0.307 0.360 0.366 0.780 0.141 0.489 0.346
DenseNet169 30 0.982 0.982 0.982 0.995 0.977 0.988 0.008
20 0.984 0.984 0.984 0.995 0.979 0.989 0.008
Classical Aug 0.505 0.548 0.637 0.856 0.421 0.635 0.219

TABLE V: Different Metrics Results

similar MSE values. We also observe that FelAug and WaterAug with all four Deep Learning models (Table IV). All performance
techniques add new image features. Comparing this with the high metrics show similar or better results with 20% augmented image
number of segments produced in the images augmented by the dataset compared to 30% augmented image dataset.
FelAug and WaterAug techniques, we can conclude that although the Our data augmentation model can be applied to various medical
superpixel augmented images are structurally similar to the original images to reduce false positive and false negative predictions while
image, the images produced by FelAug and WaterAug techniques can maintaining a high accuracy rate and low computational costs.
create more feature maps in Deep Neural networks. Hence, training
by the FelAug and WaterAug augmented image datasets provides R EFERENCES
more robust training for Deep Learning Neural Networks. Adding
new features also explains why the network produces low False [1] M. Tsuneki, ”Deep learning models in medical image analysis,” Journal
of Oral Biosciences, vol. 64, no. 3, pp. 312–320, 2022.
Positive (FPs) and False Negative (FNs) when trained with images
[2] M. H. M. Van der Velden, H. J. Kuijf, K. G. A. Gilhuijs, and M. A.
augmented by FelAug and WaterAug methods compared to classical Viergever, ”Explainable artificial intelligence (XAI) in deep learning-
augmentation methods like rotation, flip and others. based medical image analysis,” Medical Image Analysis, vol. 79, p.
102470, 2022.
Although both 20% and 30% augmented image datasets show very [3] K. Ghosh, C. Bellinger, R. Corizzo, P. Branco, B. Krawczyk, and N.
similar results, the 20% slightly outperforms the 30% augmented Japkowicz, ”The class imbalance problem in deep learning,” Machine
dataset. The 20% augmented image dataset shows lower FP and FN Learning, vol. 113, no. 7, pp. 4845–4901, 2024.
AUTHOR et al.: PREPARATION OF BRIEF PAPERS FOR IEEE TRANSACTIONS AND JOURNALS (DECEMBER 2023) 9

[4] L. Taylor and G. Nitschke, ”Improving deep learning with generic Transactions on Circuits and Systems for Video Technology, vol. 31,
data augmentation,” in 2018 IEEE Symposium Series on Computational no. 6, pp. 2114–2124, 2020.
Intelligence (SSCI), 2018, pp. 1542–1547. [26] X. Ren and J. Malik, ”Learning a classification model for segmentation,”
[5] Tensorflow. [Online]. Available: https://www.tensorflow.org/ in Proceedings of the Ninth IEEE International Conference on Computer
api_docs/python/tf/keras/preprocessing/image/ Vision, 2003, pp. 10–17.
ImageDataGenerator. [Accessed: Nov. 4, 2024]. [27] M.-Y. Liu, O. Tuzel, S. Ramalingam, and R. Chellappa, ”Entropy rate
[6] Ren and Malik, ”Learning a classification model for segmentation,” in superpixel segmentation,” in CVPR 2011, 2011, pp. 2097–2104.
Proceedings Ninth IEEE International Conference on Computer Vision, [28] J. E. Vargas-Muñoz, A. S. Chowdhury, E. B. Alexandre, F. L. Galvão,
2003, pp. 10–17. P. A. V. Miranda, and A. X. Falcão, ”An iterative spanning forest
[7] I. B. Barcelos, F. Belém, P. Miranda, A. X. Falcão, Z. K. G. do Patrocı́nio framework for superpixel segmentation,” IEEE Transactions on Image
Jr., and S. J. F. Guimarães, ”Towards interactive image segmentation by Processing, vol. 28, no. 7, pp. 3477–3489, 2019.
dynamic and iterative spanning forest,” in International Conference on [29] P. F. Felzenszwalb and D. P. Huttenlocher, ”Efficient graph-based image
Discrete Geometry and Mathematical Morphology, Springer, 2021, pp. segmentation,” International Journal of Computer Vision, vol. 59, pp.
351–364. 167–181, 2004.
[8] Y. Liang, Y. Zhang, Y. Wu, S. Tu, C. Liu, ”Robust video object
[30] E. Dougherty, Mathematical Morphology in Image Processing, vol. 1,
segmentation via propagating seams and matching superpixels,” IEEE
CRC Press, 2018.
Access, vol. 8, pp. 53766–53776, 2020.
[31] L. Vincent and P. Soille, ”Watersheds in digital spaces: an efficient
[9] B. Sheng, P. Li, S. Mo, H. Li, X. Hou, Q. Wu, J. Qin, R. Fang, and D. D.
algorithm based on immersion simulations,” IEEE Transactions on
Feng, ”Retinal vessel segmentation using minimum spanning superpixel
Pattern Analysis and Machine Intelligence, vol. 13, no. 6, pp. 583–598,
tree detector,” IEEE Transactions on Cybernetics, vol. 49, no. 7, pp.
1991.
2707–2719, 2018.
[10] L. Ren, L. Zhao, and Y. Wang, ”A superpixel-based dual window RX for [32] J. Rogowska, K. Batchelder, G. S. Gazelle, E. F. Halpern, W. Connor, and
hyperspectral anomaly detection,” IEEE Geoscience and Remote Sensing G. L. Wolf, ”Quantitative CT lymphography: evaluation of selected two-
Letters, vol. 17, no. 7, pp. 1233–1237, 2019. dimensional techniques for computed tomography quantitation of lymph
[11] W. Zhao, Y. Fu, X. Wei, and H. Wang, ”An improved image semantic nodes,” Investigative Radiology, vol. 31, no. 3, pp. 138–145, 1996.
segmentation method based on superpixels and conditional random [33] Z. Hu, Q. Zou, and Q. Li, ”Watershed superpixel,” in 2015 IEEE
fields,” Applied Sciences, vol. 8, no. 5, p. 837, 2018. International Conference on Image Processing (ICIP), 2015, pp. 349–
[12] L. Fang, S. Li, X. Kang, and J. A. Benediktsson, ”Spectral–spatial clas- 353.
sification of hyperspectral images with a superpixel-based discriminative [34] A. R. Khan, S. Khan, M. Harouni, R. Abbasi, S. Iqbal, and Z. Mehmood,
sparse model,” IEEE Transactions on Geoscience and Remote Sensing, ”Brain tumor segmentation using K-means clustering and deep learning
vol. 53, no. 8, pp. 4186–4201, 2015. with synthetic data augmentation for classification,” Microscopy Re-
[13] J.-H. Bae, G.-H. Yu, J.-H. Lee, D. T. Vu, L. H. Anh, H.-G. Kim, and J.-Y. search and Technique, vol. 84, no. 7, pp. 1389–1399, 2021.
Kim, ”Superpixel image classification with graph convolutional neural [35] B. Dufumier, P. Gori, I. Battaglia, J. Victor, A. Grigis, and E. Duch-
networks based on learnable positional embedding,” Applied Sciences, esnay, ”Benchmarking CNN on 3D anatomical brain MRI: architec-
vol. 12, no. 18, p. 9176, 2022. [Online]. Available: https://doi. tures, data augmentation and deep ensemble learning,” arXiv preprint
org/10.3390/app12189176 arXiv:2106.01132, 2021.
[14] F. Yang, Z. Ma, and M. Xie, ”Image classification with superpixels and [36] F. Isensee, P. F. Jäger, P. M. Full, P. Vollmuth, and K. H. Maier-
feature fusion method,” Journal of Electronic Science and Technology, Hein, ”nnU-Net for brain tumor segmentation,” in BrainLesion: Glioma,
vol. 19, no. 1, p. 100096, 2021. [Online]. Available: https://doi. Multiple Sclerosis, Stroke and Traumatic Brain Injuries, Lima, Peru,
org/10.1016/j.jnlest.2021.100096 2020, pp. 118–132.
[15] J. Zhang, J. Chen, Q. Wang, and S. Chen, ”Spatiotemporal saliency [37] L. Fidon, S. Ourselin, and T. Vercauteren, ”Generalized Wasserstein dice
detection based on maximum consistency superpixels merging for video score, distributionally robust deep learning, and ranger for brain tumor
analysis,” IEEE Transactions on Industrial Informatics, vol. 16, no. 1, segmentation: BraTS 2020 challenge,” in BrainLesion: Glioma, Multiple
pp. 606–614, 2019. Sclerosis, Stroke and Traumatic Brain Injuries, Lima, Peru, 2020, pp.
[16] Z. Liu, X. Zhang, S. Luo, and O. Le Meur, ”Superpixel-based spatiotem- 200–214.
poral saliency detection,” IEEE Transactions on Circuits and Systems for [38] T. Kossen et al., ”Synthesizing anonymized and labeled TOF-MRA
Video Technology, vol. 24, no. 9, pp. 1522–1540, 2014. [Online]. Avail- patches for brain vessel segmentation using generative adversarial net-
able: https://doi.org/10.1109/TCSVT.2014.2308642 works,” Computers in Biology and Medicine, vol. 131, p. 104254, 2021.
[17] V. Jampani, D. Sun, M.-Y. Liu, M.-H. Yang, and J. Kautz, ”Superpixel [39] D. Vitas, M. Tomic, and M. Burul, ”Image augmentation techniques
sampling networks,” in Proceedings of the European Conference on for cascade model training,” in 2018 Zooming Innovation in Consumer
Computer Vision (ECCV), 2018, pp. 352–368. Technologies Conference (ZINC), 2018, pp. 78–83.
[18] F. Yang, Q. Sun, H. Jin, and Z. Zhou, ”Superpixel segmentation [40] Y. Zhang et al., ”SPDA: Superpixel-based data augmentation for biomed-
with fully convolutional networks,” in Proceedings of the IEEE/CVF ical image segmentation,” arXiv preprint arXiv:1903.00035, 2019.
Conference on Computer Vision and Pattern Recognition, 2020, pp.
[41] K. Hammoudi et al., ”SuperpixelGridMasks data augmentation: Applica-
13964–13973.
tion to precision health and other real-world data,” Journal of Healthcare
[19] C. Wu and Y. Yan, ”A survey of superpixel methods and their ap-
Informatics Research, vol. 6, no. 4, pp. 442–460, 2022.
plications,” Authorea Preprints, 2023. [Online]. Available: https:
[42] D. Sun and F. Dornaika, ”Data augmentation for deep visual recognition
//doi.org/10.22541/au.169021848.16785743/v1
using superpixel-based pairwise image fusion,” Information Fusion, p.
[20] Z. Chen, B. Guo, C. Li, and H. Liu, ”Review on superpixel generation
102308, 2024.
algorithms based on clustering,” in 2020 IEEE 3rd International Confer-
ence on Information Systems and Computer Aided Education (ICISCAE), [43] Xu, L. and M. Mohammadi, ”Brain tumor diagnosis from MRI based
2020, pp. 532–537. on MobileNetV2 optimized by contracted fox optimization algorithm,”
[21] R. Achanta, A. Shaji, K. Smith, A. Lucchi, P. Fua, and S. Süsstrunk, Heliyon, vol. 10, no. 1, 2024.
”SLIC superpixels compared to state-of-the-art superpixel methods,” [44] Z. Gao, Y. Tian, S.-C. Lin, and J. Lin, ”A CT image classification
IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. network framework for lung tumors based on pre-trained MobileNetV2
34, no. 11, pp. 2274–2282, 2012. model and transfer learning, and its application and market analysis in
[22] Z. Li and J. Chen, ”Superpixel segmentation using linear spectral the medical field,” arXiv preprint arXiv:2501.04996, 2025.
clustering,” in Proceedings of the IEEE Conference on Computer Vision [45] K. B. Nampalle, P. Singh, U. V. Narayan, and B. Raman, ”DeepMediX:
and Pattern Recognition, 2015, pp. 1356–1363. A deep learning-driven resource-efficient medical diagnosis across the
[23] J. Zhao, R. Bo, Q. Hou, M.-M. Cheng, and P. Rosin, ”FLIC: Fast linear spectrum,” arXiv preprint arXiv:2307.00324, 2023.
iterative clustering with active search,” Computational Visual Media, vol. [46] G. Manohar, G. O’Reilly, and R. O’Reilly, ”InceptionCaps: A performant
4, pp. 333–348, 2018. glaucoma classification model for data-scarce environment,” in Proc.
[24] R. Achanta and S. Süsstrunk, ”Superpixels and polygons using simple 31st Irish Conf. Artificial Intelligence and Cognitive Science (AICS),
non-iterative clustering,” in Proceedings of the IEEE Conference on 2023, IEEE.
Computer Vision and Pattern Recognition, 2017, pp. 4651–4660. [47] M. S. Akter, et al., ”Multi-class skin cancer classification architecture
[25] C. Wu, J. Zheng, Z. Feng, H. Zhang, L. Zhang, J. Cao, and Y. based on deep convolutional neural network,” in Proc. 2022 IEEE
Hong, ”Fuzzy SLIC: Fuzzy simple linear iterative clustering,” IEEE International Conference on Big Data (Big Data), 2022, pp. IEEE.
10 IEEE/ACM TRANSACTIONS ON COMPUTATIONAL BIOLOGY AND BIOINFORMATICS, VOL. XX, NO. XX, XXXX 2023

[48] Z. Zhong, et al., ”Cancer image classification based on DenseNet model,” Prof. Bhaskar Mazumder (M. Pharm, PhD,
Journal of Physics: Conference Series, vol. 1651, no. 1, 2020, IOP FIC.) is currently working in Department of phar-
Publishing. maceutical Sciences, Dibrugarh University. He
[49] Y. Jiménez Gaona, et al., ”DenseNet for breast tumor classification in has pursued his B. Pharm, M. Pharm, and PhD
mammographic images,” in Proc. International Conference on Bioengi- from Jadavpur University, Kolkata. He has done
neering and Biomedical Signal and Image Processing, Cham: Springer his Post-Doctoral Fellowship in 2012 as an over-
International Publishing, 2021, pp. [include page numbers if available]. seas associate in USF College of Pharmacy,
[50] S. Liu, ”Enhancing breast cancer classification using transfer University of South Florida (Health), Florida,
ResNet with lightweight attention mechanism,” arXiv preprint USA. Prof. Mazumder has joined the department
arXiv:2308.13150, 2023. in 1999 as a lecturer and currently serving as
a Professor in the field of Pharmaceutics. He
has been the principal investigator in different projects sponsored by
AICTE, UGC, ICMR, DST, DBT, etc. Currently, he holds a German patent
for developing lipid nanoparticles. Several research and review articles
have been published in high-impact factor journals by him. He has
also published several book chapters. He has his books also that help
in understanding pharmaceutics more deeply. Prof. Mazumder is also
holding some esteemed positions in other universities. He is also noted
for his impactful communication and was invited to be a key speaker at
different esteemed universities. Till date, Prof. Mazumder has achieved
numerous accolades, and his knowledge also inspires many students to
Subhendu Sinha Chaudhuri did his B.Pharm pursue pharmacy as a profession in their life.
from Jadavpur University in 1995, and M.Pharm
in 1998. He has worked in various positions
with various multinational pharmaceutical com-
panies. He left his job in 2013 to work as a free-
lance pharmaceutical consultant. His research
interest includes, disease detection and pre-
scription management using AI/DL. Currently,
he is pursuing his PhD degree from Dibrugarh
University.

Samiran Basak (Member, IE, India) received

the AMIE degree in Electronics and Commu-
nication from Institute of Engineers (India) in
2011 and the M.Tech. degree in VLSI and Mi-
croelectronics Technology from Jadavpur Uni-
versity in 2015. He is currently a Ph.D. candidate
at Jadavpur University. His research interests
include Machine Learning Application in Bio-
Medical domain, IoT application in Agriculture.
He is presently associated as Assistant Profes-
sor with Future Institute of Technology, Kolkata.

Prof. Sheli Sinha Chaudhuri (Member, IEEE)

has been serving the ETCE Department of Ja-
davpur University for the last 26 years. After
completing her doctoral studies in 2007 from
the same department on robotic vision and path
planning, she had been actively involved in su-
pervising students for their postgraduate and
doctoral studies in the domain of Computer Vi-
sion and Machine Learning. She has supervised
sponsored projects and is currently engaged in
research collaborations with Paris-Saclay Uni-
versity, France.