Articles

Explainable Pneumonia Detection in Chest X-Rays: A Comparative Study of CNNs and Vision Transformers

/0 Comments/in /by

Pneumonia is a leading cause of global mortality, especially among children and the elderly, and chest radiography (CXR) remains the most widely used modality for its diagnosis. While deep learning has reached or exceeded radiologist-level performance on this task, the resulting models are still treated as opaque black boxes, which is a critical barrier to clinical deployment. In this work, we present a comparative and interpretable computer-aided-diagnosis (CAD) framework for pneumonia detection that combines three modern image-recognition backbones—a convolutional ResNet-50, a Swin Transformer (Swin-T), and a modernised convolutional network (ConvNeXt-T)—with Gradient-weighted Class Activation Mapping++ (Grad-CAM++) explanations. The three backbones were fine-tuned on the public Kermany chest X-ray dataset using a class-balanced training subset, weighted cross-entropy and an early-stopping protocol, and then evaluated on the held-out test set of 624 images. The Swin-T backbone achieved the best overall performance with a test accuracy of 95.51%, an F1-score of 0.95 and only 11 false negatives out of 234 normal cases, outperforming both ResNet-50 (93.11%) and ConvNeXt-T (88.94%). Grad-CAM++ heatmaps generated from the convolutional and transformer feature maps consistently localised on the affected pulmonary regions, providing radiologically plausible visual evidence for each prediction. Compared with five recent state-of-the-art pneumonia detectors, our Swin-T-based pipeline reaches a competitive accuracy while delivering layer-faithful visual explanations, supporting its use as a transparent decision-support tool in clinical workflows.

FundusSSM: A Hybrid CNN–State Space Model with Geometry-Aware Ring-Scan Tokenization for Retinal Disease Classification

/0 Comments/in /by

Automated retinal disease classification from colour fundus photographs is a critical screening tool for early diagnosis of sight-threatening conditions, especially in regions with limited access to ophthalmologists. Convolutional neural networks (CNNs) and vision transformers have achieved strong performance in this task; however, both families treat the fundus image as a generic two-dimensional grid and ignore the well-known circular geometry of fundus photography and the concentric anatomical organisation of the retina. In this paper, we propose FundusSSM, a hybrid architecture that combines a pretrained ConvNeXt-Tiny feature extractor with a geometry-aware Ring-Scan State Space Model. The Ring-Scan tokenizer partitions the CNN feature map into  equal-area concentric rings that align with the optic disc, the macula, and the peripheral retina; each ring is then processed by a bidirectional Mamba block, and information is exchanged across rings every two layers through a lightweight cross-ring attention module. We evaluate FundusSSM on a 4,217-image, four-class fundus dataset (cataract, diabetic retinopathy, glaucoma and normal) under stratified five-fold cross-validation. FundusSSM achieves the highest mean F1-score among the evaluated models (95.78%), with a low cross-fold standard deviation of 0.59% that is smaller than those of the closest baselines (ConvNeXt-Tiny and Swin-Tiny), and it outperforms ConvNeXt-Tiny, Swin-Tiny, EfficientNet-B4 and ResNet-50 in mean F1. An ablation study confirms that the proposed Ring-Scan ordering reduces the cross-fold variance by approximately 46% relative to a raster-scan ablation that uses the same architecture but a standard row-major token order. We further introduce a ring-level explainability analysis that produces per-ring feature-contribution scores aligned with clinical anatomical zones, and we observe that the model concentrates most on central optic-disc tokens for glaucoma while activating all rings nearly uniformly for cataract — patterns that agree with how clinicians read the same images. We believe that the approach followed in this research and the achieved findings could be useful to other researchers who are interested in geometry-aware deep-learning models for fundus screening tasks.

Building Trust in Agentic AI: TRACE Framework for Policy-Driven Multi-Agent System Design

/in /by

The rapid adoption of multi-agent AI systems— ranging from prescriptive, workflow-driven deployments to fully agentic, autonomous ecosystems—raises urgent challenges for trust, accountability, and regulatory compliance. This paper introduces the TRACE Framework (Trust, Review, Accountability, Critique, Explainability), a governance-first architecture designed to make multi-agent AI systems auditable, policy-aligned, and operationally reliable across varying degrees of agent autonomy. TRACE embeds governance anchors at the agent level, enforces data privacy and policy checks, supplies a dedicated Critic agent for meta-validation, and preserves human-in- the-loop oversight where required. We present a layered architecture that separates Governance & Compliance, Operational Agents, and Oversight & Assurance, and provide a concrete methodology for instrumenting agent behaviour with provenance, explainability outputs, and per-agent metrics. A formal scoring rubric—comprising agent operational metrics, critic checks, and aggregation rules—yields an Overall System Confidence (OSC) that drives automated actions, human escalation, and continuous learning. Finally, we propose a suite of operational KPIs for each layer as Governance and Compliance Indicators (GCI), Agentic Performance Metrics (APM), and Assurance Indicators (AI) that enable financial institutions and other regulated organisations to deploy multi-agent systems that are efficient, auditable, and compliant. TRACE bridges the gap between regulatory expectations and system engineering practice— providing a practical roadmap for trustworthy multi-agent AI deployment in high-stakes domains.