Topics of coursework — Master – Department of Software Systems and Technologies

List of topical topics Topics of coursework — Master. Topics are grouped by supervisors. The student can choose the proposed topic or agree his own with the supervisor.

Bychkov Oleksii Serhiiovych

10 topics

Analysis and prototyping of explainability techniques (Xai) for medical diagnostic models

Implementation and comparison of LIME, shap, Grad-CAM, Integrated Gradients on the Dr.Case model. Report with recommendations for embedding explanations in the doctor's interface.

XAI, interpretability

Analysis of the literature on multi-agent reinforcement learning and selection of an algorithm for a master's project

Review of QMIX, MADDPG, MAPPO, IPPO, VDN, Coma. Comparison table by computational complexity, scalability, learning center. Justification of selection for the specific task of coordinating agents.

LITERATURE REVIEW

Analysis of the literature and reproduction of the som + Focal-Loss NN architecture of the Dr.Case system on its own dataset

Independent reproduction of the TwoBranchNN_BMU_Focal architecture on a synthetic medical building. Benchmark vs. simpler logistic regression and CatBoost. Report with analysis of strengths and weaknesses of the approach.

Reproducibility, Dr.Case

Patent landscape analysis of medical diagnostic DSS and intelligent agent systems

A systematic review of USPTO, EPO, WIPO patents over the past 5 years on the topic of medical CDSs. Identification of "white spots" — areas with little patent activity, suitable for further work.

Patent landscape

FocusedQuestionSelector vs Basic Question Engine Benchmark on Synthetic Medical Building

Creating a corpus of 1000 synthetic diagnostic sessions. Comparison of the average number of questions before the correct diagnosis. Analysis of the impact of symptom specificity.

Active learning, benchmarking

Experimental evaluation of SymptomDrivenDiscovery on artificially "hidden" diagnoses

Preparation of a test set in which key symptoms are deliberately excluded from the training dataset. Measurement of the ability of the v2.0 module to detect such diagnoses due to pathognomonic symptoms.

Rare disease, Dr.Case

Experimental study of Focal Loss vs Cross-Entropy for unbalanced medical classes

Systematic comparison of loss functions at different levels of imbalance (1:10, 1:100, 1:1000). Selection of hyperparameters γ and α. Training schedules, ROC, PR curves.

Loss functions, imbalance

Experimental study of entropy metrics in Question Engine with different distributions of hypotheses

Comparison of Shannon entropy, Renyi entropy, Tsallis entropy as the basis of Expected Information Gain. Analysis of behavior in "flat" and "acute" distributions.

Information theory

Prototype pipeline for reprocessing radiographs for further study in the master's thesis

Implementation of the standard pipeline: histogram equalization, lung segmentation (UNet), data augmentation. Preparation of a dataset with 1000+ images from public sources.

Image preprocessing

Prototype of extracting medical terms from electronic medical records (Medical NER)

Fine-tuning XLM-RoBERTa or UkrLaBSE on a small corpus of Ukrainian-language medical records. Metrics precision/recall/F1 on named entities (symptom, drug, diagnosis, dosage).

NER, Ukrainian medical