AI-Powered Pathology Assistant - WBC Analyzer Changelog

CHANGELOG – Final Report Updates

Part – 1

Newly Added Features and Components

Inference Time Domain Adaptation Pipeline — Part 3.5

• A three-stage adaptation pipeline has been added that can be implemented without the need to retrain the model. - Steps: Reinhard color normalization; Binary routing (restriction from 5 classes to Lymphocyte/Neutrophil space); Mild Test-Time Augmentation (TTA).
• A net improvement of +31.19 points was achieved in TestB.

Shortcut Learning Resistant Training Architecture — Part 3.6

• A new training strategy was developed with the train_shortcut_resistant.py script. - 20% Foreground Cropping and 15% Background Noise Injection were applied to the images. - An XAIFocusMonitor callback class was added that performs autonomous focus measurement via Grad-CAM at the end of the epoch.

Background Segmentation

• OpenCV-based background masking (Otsu Thresholding + Morphological Operations) integrated into the shortcut-resistant training architecture.

---

Ablation Analysis and Comparative Experiments

Summary Results

• Baseline (TestB): 56.96%
• Binary routing: 73.90% (+16.94 pp)
• Reinhard: 86.46% (+12.56 pp)
• TTA (final): 88.15% (+1.69 pp)

Experiment Comparison Table

Configuration	TestB Accuracy	Change
Baseline	56.96%	—
Binary routing	73.90%	+16.94 pp
Reinhard	86.46%	+12.56 pp
TTA (final)	88.15%	+1.69 pp

---

Changed and Updated Content

Test Set Accuracy

• TestA accuracy: 93.4% → 98.53% (+5.13 points)
• TestB accuracy (overall): 89% → 88.15%
• TestB accuracy (baseline → adaptation): 56.96% → 88.15%

Classification Report Updates

• Neutrophil Precision (TestA): 0.987 → 0.9958 (+0.009)
• Weighted avg F1 (TestA): 0.956 → 0.9853 (+0.029)

Metric	Before	After	Change
Basophil Precision	0.937	1.0000	+0.063
Monocyte F1	0.866	0.9376	+0.0716
Eosinophil F1	0.849	0.9541	+0.1051
Lymphocyte F1	0.964	0.9865	+0.0225

Confidence Score and Response Time

• Neutrophil test confidence score: 99.8% → 97.7%
• Response time: ~200 ms (unchanged)

Preprocessing Flowchart updated

---

Domain Shift Analysis and Constraints

• It was shown that the largest source of domain shift is staining/color variation and that this is corrected with Reinhard normalization. - It was noted that the Raabin-WBC dataset only includes Giemsa staining and that full-size WSI images are not yet supported.

---

Discussion Section Expansions

The discussion section in the final report was divided into 7 subheadings:

• Literature Comparison
• Class Imbalance
• Shortcut Learning and XAI
• Domain Shift and TTA
• Clinical MLOps
• Original Contributions
• Limitations

---

Reference Updates

• References: 27 → 43

• Added fields: Domain shift, TTA, Shortcut learning (Panboonyuen 2026, Geirhos 2020, Bassi 2024), Clinical MLOps (Spadacini 2026, Ali 2026), LLM-based XAI control (Mermigkis 2026), Stain-aware domain alignment (Li 2026)

---

Abstract Updates

Key points added to the abstract:

• Domain shift management

• TestB baseline 56.96% → after adaptation 88.15%
• Shortcut-Resistant Training
• +31.19 points improvement

---

New Sections Added to the Final Report

• § 3.5 — Domain Shift Management and Inference Time Improvements
• § 3.6 — Shortcut-Resistant Training Architecture
• § 4.1.2 — Inter-Experimental Comparative Analysis and Shortcut Learning Finding
• § 4.1.3 — Inference-Time Domain Adaptation and Ablation Analysis
• § 5.3–5.7 — Discussion subsections (XAI, domain shift, MLOps, contributions, limitations)

Part – 2

Model and Provider Configuration

• The primary model was configured as openai/gpt-4o via GitHub Models (with GITHUB_TOKEN).
• The secondary (fallback) provider was retained as Gemini 2.5-flash.
• The backend now returns the actual provider used for each response.
• The model badge (agentReportModelBadge) in the frontend has been made dynamic and shows the actual provider.

Token and Environment Variable Management

• The old MODEL_TOKEN has been completely removed.
• A normalized, secure read mechanism has been added for GITHUB_TOKEN and the optional GEMINI_API_KEY with the get_env_token function.

Multimodal and Fallback Logic

• Multimodal calls are wrapped with try/except. - Automatic text-only fallback is applied in case of visual support rejection or error.
• Response parsing process strengthened (robust parsing, extract_completion_text).

Model Parameters and Error Handling

• GitHub/OpenAI model parameters updated (e.g., max_tokens compliance ensured).
• Parameter mismatches leading to 400 series errors fixed.

Response Reliability and Backup Mechanism

• Response parsing mechanism made more robust.
• A deterministic, rule-based final fallback layer (build_rule_based_report) capable of producing a result in every case has been added.

Prompt and Style Improvements

• Global SYSTEM_INSTRUCTION updated. - Class-weighted CLASS_MORPHOLOGY_CONTEXT added.
• A dynamic build_agent_prompt system has been implemented to reduce repetitive and template-like reports.

API Call Ordering

• Call order has been reorganized:
  1. GitHub (GPT-4o)
  2. Gemini fallback in case of failure
• The backend returns the provider used for each response and passes it to the frontend.

Frontend Updates

• The model badge on index.html has been made dynamic. - The model information shown to the user is now updated according to the actual provider from the backend.

Documentation Updates

• README.md and README.tr.md have been updated. - Required environment variables (GITHUB_TOKEN, GEMINI_API_KEY) and model ordering have been documented.

Modified Files

• app.py

• index.html

• README.md

• README.tr.md

Test / Validation Status

• Static and syntax checks were performed; no errors were found. - End-to-end validation with actual API keys is pending in the user environment.