Eclipse: a feedback-driven framework and automated quality assessment system for reliable, interpretable, and multi-modal biomedical AI

This dissertation presents ECLIPSE (Evaluation and Clinical Pipeline for Intelligent Systems in Cancer Engineering), an overarching framework for reliable, interpretable, and resource-aware biomedical AI. ECLIPSE comprises two main components: (i) the CancerAI Workflow Assistant, a stage-gated web application that structures the AI development lifecycle into four coupled stages—Planning, Data, Modeling, and Inference—with measurable acceptance bands and closed-loop feedback; and (ii) MERIT-C (Medical Evidence and Reporting Index for Technical Cancer AI), the quality-assessment subsystem embedded within ECLIPSE that operationalizes the same four stages as a standardized scoring rubric for evaluating published cancer AI research. MERIT-C combines a Reporting Quality Score (RQS, 59 points across the four ECLIPSE stages) with a Technical AI Merit Score (TAMS, seven dimensions: Novelty, Architecture Complexity, Parameter Efficiency, Hardware Resource Characterization, Scalability, Generalizability, and Adaptability) under the composite MERIT-C = α · RQSₙₒᵣₘ + β · TAMSₙₒᵣₘ. Validated against 89 peer-review scores from OpenReview (Spearman ρ = 0.346, p = 0.015; AUC-ROC = 0.852, p = 0.001), MERIT-C achieves Good-class discrimination between accepted and rejected cancer AI papers. Applied to the dissertation’s own published papers, MERIT-C yields composite scores of 62.4–64.2 for the oral cancer WSI papers (OCU-Net, AUSAM, OCANet/LGA), with strongest performance on Architecture Complexity (75.0%) and weakest on Hardware Reporting (33.3%); 49.7–58.8 for the EEG papers (DL-IoT, ETSNet), with notable strength in Adaptability (87.5%) and a gap in Inference-stage reporting (27.8–44.4%); and 44.1–57.7 for the MRI/CT papers (Auto Claustrum, CU-Net, AM-UNet, TLU-Net), lagging the field on Novelty (−24.6 pp) and Hardware Reporting (−24.4 pp). The CancerAI Workflow Assistant is instantiated across three modality-specific case studies. Histopathology (oral cancer): OCANet achieves Dice 86.14 on ORCA and 94.09 on OCDC; AUSAM raises case-level Dice from 23.72 to 87.59 (ORCA) and 91.84 (OCDC). EEG (psychiatric/BCI): task-matched 1D CNNs reach 91.25% accuracy; fusion of activity-specific experts achieves 99.57% multi-class accuracy; subject-specific transfer improves from 72.82% to 93.33%. MRI/CT: CU-Net achieves liver Dice 0.894; the 0.34M-parameter AM-UNet attains Dice 0.82 and ICC = 0.90; TLU-Net demonstrates consistent cross-dataset gains via transfer learning. Together, the two components of ECLIPSE close complementary gaps: the Workflow Assistant provides the principled development pipeline, while MERIT-C provides the externally validated quality instrument, collectively narrowing the distance between algorithmic performance and clinically deployable, auditable AI.

Name thesis:degree_name

Ph.D. (Doctor of Philosophy)

Level thesis:degree_level

Doctoral

Discipline thesis:degree_discipline

Computer Science (UMKC)

Grantor

University of Missouri--Kansas City

Year dc:date.issued

2026

Author dc:creator

• Shah, Syed Jawad Hussain

Advisor dc:contributor.advisor

• Lee, Yugyung, 1960-

Language dc:language.iso

en_US

Handle dc:identifier.uri

https://hdl.handle.net/10355/112323

OAI identifier oai:identifier

oai:mospace.umsystem.edu:10355/112323