University of Hyderabad · 2026

Advances in Deep Learning for Medical Image Segmentation and Classification

My doctoral research tackled accuracy, computational cost and generalisation in medical image analysis, with breast ultrasound and tumor detection as the proving ground. Advisor: Dr. Avatharam Ganivada, School of Computer and Information Sciences. Funded by the UGC Junior Research Fellowship.

Key Contributions

EfficientU-Net: parameter-optimized segmentation

EfficientNet-B7 and atrous convolution folded into U-Net, tuning breast tumor segmentation and classification in ultrasound for lower computational cost.

13× reduction in parameters — 1.31M against U-Net's 17.27M
Stronger segmentation of malignant tumors with irregular shapes
Better boundary localisation through adaptive receptive fields
97.905% accuracy classifying benign, malignant and normal tissue
Validated on two public datasets with 5-fold cross-validation

UMA-Net with adaptive loss functions

A U-Net variant with residual connections, attention blocks and atrous convolutions, trained under a dynamic ensemble loss.

Residual connections and attention blocks sharpen feature integration
Atrous convolutions capture multi-scale context without losing resolution
Dynamic ensemble loss (BCE, Dice, Hausdorff, Tversky) rebalances weights during training
Generalises across five breast ultrasound datasets: BUET, BUSI, Mendeley, OMI, UDIAT

Fuzzy rough set loss for boundary precision

A loss function built on fuzzy rough set theory to handle boundary uncertainty in medical images.

Higher sensitivity to uncertain predictions and ambiguous lesion boundaries
Lower computational complexity alongside improved segmentation accuracy
Handles irregular shapes and overlapping edges
New similarity functions for uncertainty in boundary regions

Deep learning and genetic algorithm ensemble

MobileNet for feature extraction, genetic algorithms for feature selection, and an ensemble classifier with soft voting.

MobileNet as a minimal-parameter feature extractor for medical images
GA-based selection navigating complex feature spaces
Soft-voting ensemble for robust classification decisions
Addresses overfitting where data is limited

Saliency-Guided AttentionNet (SGAN)

A dual-branch architecture using Grad-CAM saliency maps for breast ultrasound classification.

90.51% accuracy on multi-center validation
87.95% F1-score and 94.08% AUC across five datasets
Explicit foreground–background decomposition for lesion and peritumoral analysis
Adaptive attention fusion, keeping transfer learning benefits at minimal parameter cost

Impact & Validation

The methods were evaluated on public datasets:

Five public breast ultrasound datasets (BUET, BUSI, Mendeley, OMI, UDIAT)
The fuzzy rough set loss was also tested on polyp, brain MRI, chest CT and skin-lesion datasets, with statistically significant gains (p < 0.001)
Cross-dataset testing: SGAN reached 78.46% accuracy on held-out datasets, against 90.51% on its multi-center evaluation
Compact models: EfficientU-Net uses 1.31M parameters against U-Net's 17.27M

Throughout, the emphasis is on methods that stay accurate with few parameters and little data — a prerequisite for use in health systems with limited access to expert radiologists, though clinical evaluation is still future work. Each method is published in a peer-reviewed venue, with code released publicly.

Back to research