Artificial Intelligence, Computer Vision and Pattern Recognition
30
Scopus Publications
Scopus Publications
Explainable Transformer-Augmented U-Net for Brain Tumor Segmentation in MRI Hariharan Ramamoorthy, Fathima M. Dhilsath, V. Rajalakshmi, A. Aarya Research Advance S in Intelligent Computing Volume 3, 2026 Accurate segmentation of brain tumors from magnetic resonance imaging (MRI) scans is essential for clinical decision-making, impacting diagnosis, treatment planning, and prognosis. Despite some progress made in deep learning-based methods, key challenges remain, including noise in MRI images, class imbalance, and the lack of interpretability in model predictions. This study introduces a Transformer-Augmented U-Net designed to enhance segmentation accuracy and model transparency through advanced preprocessing, Explainable AI (XAI), and optimization techniques. The preprocessing pipeline includes Non-Local Means (NLM) filtering for noise suppression, histogram matching for intensity normalization, and wavelet-based super-resolution to retain finer structural details. To address interpretability concerns, we incorporate Grad-CAM++ for visualizing attention maps, SHAP (SHapley Additive exPlanations) for analyzing feature contributions, and Layer-wise Relevance Propagation (LRP) to improve decision transparency. Additionally, the model is optimized using Tversky Loss to mitigate class imbalance and Bayesian hyperparameter tuning for improved segmentation performance. Experiments conducted on the BraTS 2023 dataset demonstrate that the proposed approach achieves a Dice Similarity Coefficient (DSC) of 0.92 and an Intersection over Union (IoU) of 0.87, surpassing conventional U-Net and Attention U-Net baselines. Qualitative analysis through XAI-based visualizations confirms the model’s ability to precisely identify tumor regions. The proposed framework not only enhances segmentation accuracy but also provides clinically interpretable insights, making it a promising tool for real-world medical applications.
Evaluating and Evading Machine Learning Malware Detectors with Deep Q Learning Karthiga Jothi Preetha V, M.Dhilsath Fathima Proceedings of 8th International Conference on Computing Methodologies and Communication Iccmc 2025, 2025 Traditional machine learning-based malware detectors often rely on static features such as API calls, DLL imports, and PE header values. However, these systems remain vulnerable to adversarial evasion techniques that subtly modify malware without altering its functionality. In this work, we address the critical issue of classifier vulnerability by proposing a reinforcement learning-based adversarial evasion framework. A Random Forest classifier is first trained on a structured dataset of labeled malware and benign samples. To assess and exploit weaknesses in the classifier, we design a Deep Q-Learning (DQL) agent integrated into a custom OpenAI Gym environment that simulates adversarial behavior. The agent iteratively modifies binary feature vectors of malware samples, guided by a dynamic reward function based on the classifier’s confidence scores, aiming to achieve misclassification with minimal feature changes. Our results show that the DQL agent achieves a high evasion success rate, significantly reducing classifier accuracy on perturbed samples, while preserving malware behavior. The proposed framework highlights the need for robust defense mechanisms in static malware detection and serves as a testbed for evaluating classifier resilience against adaptive adversaries.
Deep - Transfer Learning for Multi-Crop Leaf Disease prediction using ResNet and ConvNet M. Dhilsath Fathima, Akash Gupta, Kartik Jain International Conference on Advancements in Power Communication and Intelligent Systems Apci 2024, 2024 Plant diseases can be categorized based on the nature of their primary causal agents, either infectious or noninfectious. The detection and identification of such plant diseases play a critical role in sustaining agricultural productivity and preventing crop losses. Manual disease tracking is a labor-intensive, knowledge-dependent, and time-consuming task. To address these challenges, Deep-TL (Deep Transfer Learning) techniques are used to capture plant leaf images and compare them with extensive datasets representing various plants. This paper suggests a model that makes use of the capabilities of deep learning algorithms, particularly those trained on large datasets, such as ResNet (Residual Networks) and ConvNet (Convolutional Neural Networks). Through the use of transfer learning strategies, the Deep-TL model gains better accuracy and performance by utilising the knowledge gained and achieved the accuracy of 98.2%.
Air Quality Prediction using Deep Learning models M. Dhilsath Fathima, Sashank Donavalli, Harshitha Kambham International Conference on Advancements in Power Communication and Intelligent Systems Apci 2024, 2024 In today's world, air pollution is an important concern for both the general public and the environment. Monitoring and forecasting pollutant concentrations in the environment is critical to effective pollution control and management. Using advanced deep-learning techniques, this study focuses on the development of a time series air quality prediction (AQP) model for numerous pollutants in the air, including small particles of PM2.5, PM10, ground-level ozone (O3), carbon monoxide (CO), sulphur dioxide (SO2), and nitrogen dioxide (NO2). Time series data are used in the study to capture temporal relationships and seasonal trends inherent in air quality measurements. This proposed AQP model evaluates many deep learning architectures, such as Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Bidirectional LSTM (Bi- LSTM). This study demonstrates how effective these models are at capturing temporal dependencies and seasonal patterns in air quality datasets, pointing towards novel approaches for precise and timely air quality monitoring. The results of this research provide a substantial contribution to the advancement of air quality prediction systems, allowing for better decision- making for public health and environmental protection.
ACAC: Automatic Cardiac Arrhythmia Classification Based on 1D-Deep Resnet M. Dhilsath Fathima, S Sithsabesan, Jayanthi K, Dhanyaa S U Proceedings of Inc4 2024 2024 IEEE International Conference on Contemporary Computing and Communications, 2024 Arrhythmias are infrequent irregularities in a person's heartbeat rhythm. They result from any disruption in the ordered pattern of the heart's expanding wave of excitation, a process heavily reliant on electrical activity. These arrhythmias might result in consequences that could be fatal and constitute a direct risk to life. As a result, detecting and classifying arrhythmias is a significant challenge in cardiac diagnostics. Electrocardiograms (ECG) are widely used in the diagnosis of arrhythmias because they are a low-cost, non-invasive, and rapid technique of diagnosis. However, the unpredictability of arrhythmic segments, along with the sensitivity of ECGs to noise, leads to arrhythmia misdiagnosis. Furthermore, manually identifying cardiac arrhythmias using ECG data takes time and is liable for inaccuracy. Deep learning (DL) is a preferred approach for fast and automatic ECG signal classification, surpassing the performance of traditional machine learning models. The suggested study proposes a novel deep learning architecture, namely t-dimensional deep residual neural network (1D-deep ResNet) for an Automatic cardiac arrhythmia classification (ACAC) system. ECG signals from the MIT -BIH dataset were used to train and evaluate the model. This proposed model contributes in two different ways. First, the unbalanced ECG data are corrected using the upsampling approach to reduce noise and prevent biased prediction outcomes. Next, the unbalanced ECG signals are automatically classified using 1D-Deep ResNet. ACAC system aims to overcome the issues of traditional electrocardiograms (ECGs) in the identification of arrhythmias, which can be impacted by noise and unpredictability of events, leading to misinterpretation and mistakes. The confusion matrix is used to calculate the model accuracy, AUC, precision, recall, and f1 score to evaluate the model performance. The results of the experiments shows that the proposed method performed well, with 99.9% and 99.94% AUC score in the training and testing datasets, respectively. The proposed model outperforms other existing deep learning approaches like CNN, LSTM, and GRU in terms of performance, and it will significantly minimise the involvement of doctors in the classification of ECG signals. According to our research findings, the 1D-deep ResNet is more suitable for automated cardiac arrhythmia classification in the future than other deep learning algorithms now available.
Automatic Title Generation with Attention-Based LSTM M. Dhilsath Fathima, M. Seeni Syed Raviyathu Ammal, Prashant Kumar Singh, Sachi Shome, Manbha Kharsyienlieh, R. Hariharan Lecture Notes in Electrical Engineering, 2023
Handwritten digits classification through multi-classifier bag of visual words International Journal of Innovative Technology and Exploring Engineering, 2019
Diagnosis of acute myocardial infarction using random forest classifier through SPECT International Journal of Innovative Technology and Exploring Engineering, 2019
Privacy preserving multi-party hierarchical clustering over vertically partitioned dataset using semi-honest model International Journal of Pharmacy and Technology, 2016
K-means clustering algorithm to improve website performance International Journal of Applied Engineering Research, 2016
Identification of bacterial contamination: A classification approach using support vector machine International Journal of Applied Engineering Research, 2014
