Optimized deep learning for ultrasound classification and XAI segmentation

Abstract

Artificial Intelligence (AI) has shown transformative potential in medical imaging, and offers powerful tools for early diagnosis, disease monitoring, and decision support. There have been significant advancements at the convergence of artificial intelligence and medicine in recent years, particularly for medical image analysis applications. A critical knowledge gap persists in the field: most existing deep learning models function as "black boxes," lacking the interpretability required for high-stakes clinical decision-making. Despite witnessing transformative shift with the integration of deep learning (DL) and artificial intelligence (AI), and enabling unprecedented accuracy in diagnosing complex diseases, the black-box nature of these models has raised significant concerns regarding their interpretability, particularly in high-stakes healthcare applications. Furthermore, research has historically been conducted in silos, focusing on isolated downstream tasks rather than unified, adaptive frameworks capable of handling multiple imaging modalities.
To address these challenges, the development of efficient, robust and explainable AI (XAI) frameworks is the need of the hour, and has recently caught the attention of research community, with proliferation of research findings at a rapidly evolving pace, aiming to bridge the gap between AI-driven predictions and clinical decision-making.
This thesis proposes an innovative computational framework based on an Agentic AI paradigm. This framework integrates optimized deep learning models with Explainable AI (XAI) techniques into a modular, adaptive pipeline. It is designed to autonomously plan tasks—such as detection, segmentation, and localization—while providing transparent, interpretable visual evidence for its predictions.
The evaluation of the proposed framework and optimized algorithm pipeline based on deep learning and explainable AI (XAI) interpretation and visualisation is done for three different clinical downstream tasks - the retinal vasculature extraction for diabetic retinopathy detection, semantic segmentation and pathology localization in lung ultrasound (LUS), and automated thyroid nodule detection from ultrasound imaging. The extensibility and adaptive feature of the proposed framework for new medical imaging settings was also examined for a new downstream clinical task. involving pituitary brain tumour detection from MRI scans. The goal was to develop a robust, interpretable, extensible, adaptive and clinically meaningful AI system, capable of expert-level interpretation of medical imagery, built on cross-domain transfer learning, which allows adaptation of efficient foundation models, to new downstream task, and translation of their capability to solve more complex downstream tasks in clinical scenarios, based on pre-trained models using transfer learning and explainable machine learning techniques. The thesis encapsulates vital contributions and insights pertaining to each area discussed, and potential to extend it to a wide range of clinical settings across various medical fields.
Key Findings:
The evaluation of this framework across three primary clinical tasks yielded several key findings:
• Retinal Vasculature Extraction: The use of an integrated residual-attention U-Net++ significantly improved the clarity of vascular structures, outperforming traditional segmentation methods.
• Lung Ultrasound (LUS) Segmentation: Implementing DeepLabV3+ with XAI overlays provided precise semantic segmentation even under weak supervision, accurately localizing pathologies like pleural effusion and consolidation.
• Thyroid Nodule Detection: A hybrid system combining YOLOv8 with radiomics-informed reasoning achieved expert-level detection while providing clinical interpretability through SHAP and Grad-CAM++ visualizations.
• Framework Extensibility: The framework demonstrated high adaptability when extended to a new downstream task involving pituitary brain tumour detection from MRI scans, validating its potential for cross-domain clinical application.
The research concludes that an integrated, agentic approach not only matches human-level diagnostic accuracy but also bridges the trust gap between AI and clinicians. By transforming opaque models into transparent assistive tools, this framework paves the way for the ethical and practical deployment of AI in diverse, real-world healthcare settings.

Date of Award	2026
Original language	English
Supervisor	Girija CHETTY (Supervisor) & Elisa Martinez-Marroquin (Supervisor)