Work

Domain-Agnostic Fuzzy Deep Learning Framework for Explainable AI in Multimodal Diagnostics and Decision Support

Year: 2026
Type: dissertation
Abstract: Despite achieving strong predictive accuracy in medical diagnosis, the widespread adoption of deep learning-powered clinical decision support systems is hindered by a fundamental conflict between this accuracy and the need for interpretability. This project addresses that tension by designing, implementing, and evaluating the Domain-Agnostic Fuzzy Deep Learning (DFDL) Framework. This multimodal, multidomain clinical decision support system integrates the intrinsic interpretability of fuzzy logic with the representational power of deep neural image encoders. Existing deep learning models present a core problem: their operation is opaque, functioning as "black boxes." While they generate predictions, they cannot provide human-readable explanations detailing the specific clinical factors that led to a particular decision. This is unacceptable in healthcare, where clinicians require transparent reasoning to validate AI-assisted diagnoses and ensure patient safety. Moreover, no single imaging modality captures the full diagnostic picture, as effective systems must fuse volumetric imaging data with structured electronic health records (EHRs). The DFDL framework addresses both challenges through a two-stream hybrid architecture. Raw volumetric scans (DICOM, NIfTI) are passed through a frozen MedicalNet-pretrained 3D-ResNet-10 backbone to produce 512-dimensional L2-normalized feature vectors, which are stored as NPZ files for efficient reuse. These image features are processed by a flat MLP encoder (FlatEncoder). At the same time, clinical tabular data is modelled by an Adaptive Neuro-Fuzzy Inference System (ANFIS) using mixed Gaussian and trapezoidal membership functions. A learnable dynamic gate fuses both streams per-sample, automatically falling back to clinical-only inference when imaging is absent. A FastAPI backend exposes the model through structured REST endpoints, and a Retrieval-Augmented Generation (RAG) copilot powered by ChromaDB and NVIDIA NIM (Llama 3.1) translates model outputs into natural-language explanations for clinicians. The framework was evaluated on the OASIS-1 cross-sectional Alzheimer’s dataset (BAcc=0.9052±0.0116, AUC=0.9593±0.0088), OASIS-2 longitudinal Alzheimer’s dataset (BAcc=0.8465±0.0260, AUC=0.8836±0.0288), MMIST-ccRCC renal cell kidney cancer dataset (BAcc=0.6506±0.0439, AUC=0.6188±0.0563), and UTSW-Glioma molecular brain tumor dataset (BAcc=0.8353±0.0327, AUC=0.8347±0.0545). All results are produced via unbiased nested cross-validation (5-outer and 3-inner folds). The system's configuration-driven architecture allows for easy generalization to any medical domain, enabling the addition of different imaging modalities or the switching of datasets. Overall, this project demonstrates the design, implementation, and evaluation of a complete computing-based solution, from raw data ingestion to user-facing clinical explanation, addressing a real and significant gap in medical AI deployability.
Language: English
Open Access status: green
OpenAlex ID: W7198992755
Cites: 24
Cited by:
Related to: 10