Agents that Argue and Explain Classifications of Retinal Conditions

Expertise for auditing AI systems in medical domain is only now being accumulated. Conformity assessment procedures will require systems: (1) to be transparent, (2) not rely decisions solely on algorithms, or (3) include safety assurance cases the documentation facilitate technical audit. We are interested here obtaining transparency case of machine learning (ML) applied classification retina conditions. High performance metrics achieved using ML has become common practice. However, domain, algorithmic need sustained by explanations. aim at building a support tool ophthalmologists able to: (i) explain decision human agent automatically extracting rules from learned models; (ii) ophthalmologist loop formalising expert and including knowledge argumentation machinery; (iii) build creating argument patterns each diagnosis. For task, we used dataset consisting 699 OCT images: 126 Normal class, 210 with Diabetic Retinopathy (DR) 363 Age Related Macular Degeneration (AMD). The contains patients Ophthalmology Department County Emergency Hospital Cluj-Napoca. All ethical norms procedures, anonymisation, have been performed. three algorithms: tree (DT), vector (SVM) artificial neural network (ANN). algorithm extract diagnosis rules. knowledge, relied normative (Invernizzi et al. Ophthalmol Retina 2(8):808–815, 2018). arguing between agents, Jason multi-agent platform. assume different base reasoning capabilities agent. agents their own optical coherence tomography (OCT) images which they apply distinct algorithm. model With rules, engage an argumentative process. resolution debate outputs that then explained ophthalmologist, means cases. diagnosing condition, our solution deals following issues: first, models translated into These explanation tracing chain supporting Hence, proposed complies requirement “algorithmic should agent”. Second, based ML-algorithms. architecture includes knowledge. taken exchanging arguments ML-based algorithms conflict among verbalised, so can supervise Third, generated structure evidence various goals such as: methodology, transparency, data quality. dimension, auditor check provided against current best practices standards. developed system conditions goes behind most software focuses metrics. Our approach helps approve domain. Interleaving extracted ML-models step towards balancing benefits explainability, aiming engineering reliable applications.