PathoClass-BRCA: Reframing Pathology Report Generation as Guideline-Aligned Multi-Task Classification

Official implementation of PathoClass-BRCA: Reframing Pathology Report Generation as Guideline-Aligned Multi-Task Classification (BMVC 2026), a framework that reformulates pathology report generation from whole-slide images (WSIs) as a clinically grounded multi-task classification problem.

Prior approaches treat report generation as a captioning task and evaluate models with surface-form lexical metrics (BLEU, ROUGE, METEOR). We instead predict standardized pathological features defined by the ICCR reporting guidelines, enabling evaluation based on diagnostic correctness rather than linguistic similarity.

Left: label-extraction pipeline. Centre: pathological feature-level evaluation protocol. Right: classification-oriented reformulation.

---

📑 Contents

• 📑 Contents
• 📜 Overview
  • Motivation
  • Contributions
• 🛠️ Installation
• 🏥 PathoClass-BRCA Dataset
• 🔬 Data Preparation
  • 1. Downloading TCGA Slides
  • 2. Whole-Slide Image Processing
  • 3. Feature Extraction
• 🚀 Training
• 📊 Evaluation
• 🔗 Citation
• References

---

📜 Overview

Motivation

Histopathology is the gold standard for cancer diagnosis, and pathology reports are central to therapeutic planning and prognostic assessment. Recent works automate report generation from WSIs with multimodal LLMs, but most methods:

• optimise token-level cross-entropy losses designed for captioning;
• evaluate with n-gram overlap metrics;
• reward linguistic similarity rather than clinical correctness.

We reformulate report generation as a six-task, guideline-aligned classification problem, explicitly predicting the morphological features routinely assessed in invasive breast carcinoma.

Contributions

	Contribution	Where in the repo
i	PathoClass-BRCA, the first breast cancer WSI benchmark reframing report generation as multi-task classification over six ICCR-aligned features	dataset/
ii	A pathological feature-level semantic evaluation protocol mapping free-text reports to guideline-defined labels via LLM extraction, scored with macro-F1 / macro-recall	evaluation/
iii	Classification-oriented adaptations of report generation backbones using learnable task prompts and multi-head classifiers	models/*_classification.py

The six pathological features

Task (--tasks name)	Classes	Coverage
diagnosis (histological type)	Invasive ductal / unspecified, invasive lobular, other	100.0%
grade (histological grade)	Grade 1, Grade 2, Grade 3	90.2%
nuclear_pleomorphism	High, Moderate, Low/absent	94.8%
mitotic_activity	High, Moderate, Low/absent	87.1%
tubular_formation	High, Moderate, Low/absent	94.7%
LVI (lymphovascular invasion)	Present, Absent	82.7%

---

🛠️ Installation

git clone https://github.com/AImageLab-zip/PathoClassBRCA.git
cd PathoClassBRCA

conda create -n pathoclass python=3.10
conda activate pathoclass
pip install -r requirements.txt

Experiments were run with PyTorch 2.7.1 on NVIDIA L40S (48 GB) GPUs.

---

🏥 PathoClass-BRCA Dataset

We publicly release PathoClass-BRCA, a classification-oriented benchmark for breast cancer WSIs built on top of the BRCA subset of WSI-Bench. Preliminary labels are extracted from free-text reports with GPT-5 using feature-specific prompts, then reviewed and corrected by pathologists (96.7% agreement between automatic and pathologist-verified labels).

dataset/
├── classification/                           ← ★ PathoClass-BRCA (this work)
│   └── BRCA/
│       ├── feature_datasets/                 Per-feature labels
│       │   ├── diagnosis_dataset.json
│       │   ├── grade_dataset.json
│       │   ├── nuclear_pleomorphism_dataset.json
│       │   ├── mitotic_activity_dataset.json
│       │   ├── glandular_tubular_formation_dataset.json
│       │   └── vascular_lymphovascular_invasion_dataset.json
│       └── splits/                           
│           ├── train_multitask.json
│           ├── val_multitask.json
│           └── test_multitask.json
│
└── report_generation/                        Source reports, from WSI-Bench BRCA
    └── BRCA/
        ├── WSI-Bench-train-report.json
        ├── WSI-Bench-val-report.json
        ├── WSI-Bench-test-report.json
        ├── short_wsi_bench.json              Sentence corpus for BiGen retrieval
        └── split_3_wsi_bench.csv             Slide-level train/val/test partition

• classification/ — PathoClass-BRCA, our contribution: multi-task annotations mapping each slide_id to the six ICCR-aligned feature labels (-1 = not documented in the report).
• report_generation/ — the free-text reports from the BRCA subset of WSI-Bench, from which those labels are extracted.

Label distribution

Label distributions for the six selected pathological features (after pathologist review).

Several tasks are strongly imbalanced — most visibly histological type (73% invasive ductal) and tubular formation (773 low/absent vs 29 high).

---

🔬 Data Preparation

1. Downloading TCGA Slides

Diagnostic WSIs (.svs) are available from the NIH Genomic Data Commons Data Portal. TCGA-BRCA (Breast Invasive Carcinoma) slides can be retrieved with the GDC Data Transfer Tool.

2. Whole-Slide Image Processing

Tissue regions are segmented with Otsu's thresholding on a downsampled WSI via OpenSlide, and non-overlapping 256 × 256 patches are extracted at 20× magnification. We use CLAM for the whole pre-processing stage:

python CLAM/create_patches_fp.py \
    --source <SVS_DIR> \
    --save_dir <PATCH_DIR> \
    --patch_size 256 \
    --step_size 256 \
    --seg --patch --stitch

3. Feature Extraction

To extract embeddings with UNI, PLIP or GigaPath, refer to extract_embeddings/ and its own README, which describes the scripts and how to launch them in detail.

🚀 Training

All models are trained through main.py, which handles both paradigms: adding --classification switches from free-text report generation to multi-task classification — see the training guide for the full command of every model and paradigm.

📊 Evaluation

Generative models (pathological feature-level evaluation protocol)

Free-text reports produced by MiGen / HistGen / BiGen / WSI-LLaVA are mapped to the six standardized labels with the feature-specific prompts in evaluation/prompts.py, then scored against PathoClass-BRCA ground truth. Two scripts are provided, differing only in the format of the predictions they read:

• extract_labels_multitask_from_report.py to evaluate MiGen, HistGen, BiGen;
• extract_labels_multitask_from_report_wsi_llava.py to evaluate WSI-LLaVA.

python -m evaluation.extract_labels_multitask_from_report \
    --root <PREDICTIONS_DIR> \
    --gt dataset/classification/BRCA/splits/test_multitask.json \
    --model <MODEL_NAME> \
    --save_name <SAVE_NAME> \
    --model_type gpt-5 \
    --tasks all

root is a directory of .txt files, one per sample, each holding a Python dict with predict / target. --tasks is a comma-separated subset of the six task names, or all (default).

python -m evaluation.extract_labels_multitask_from_report_wsi_llava \
    --jsonl <PREDICTIONS_FILE>  `# JSONL predictions` \
    --text_field answer         `# which JSON field to score` \
    --gt dataset/classification/BRCA/splits/test_multitask.json \
    --model <MODEL_NAME>        `# label only: results/<model>/` \
    --save_name <SAVE_NAME>     `# suffix for the output files` \
    --model_type gpt-5          `# OpenAI model used for extraction` \
    --tasks all                 `# tasks to score, or a comma-separated subset`

Each run writes three things under classification/results/<model>/: a metrics_*.txt log, a preds_*.csv with one row per (sample, task), and a gpt_cache_*/ directory. Extraction results are cached per sample and task, so re-running after a crash — or with a different --tasks subset — only issues the missing API calls.

The scripts:

1. extract structured pathological features from each generated report via LLM prompting;
2. maps the extracted text back to a label using the vocabulary built from the ground truth; unmappable answers are logged as PRED_TEXT_NOT_IN_ALLOWED and dropped;
3. discard samples whose ground-truth label is -1;
4. compute macro-F1 and macro-recall against the ground truth.

If the ground truth is available and the model omits the feature, i.e. the feature label is -1, this case is counted as an error for the true class. Pathologists reviewed labels extracted from 1,500 generated reports; the extraction protocol reaches 96.4% concordance in identifying the intended feature values.

⚠️ This step queries an external LLM. Export your API key before running:

export OPENAI_API_KEY=<YOUR_KEY>

---

🔗 Citation

If you find this code or dataset useful, please cite our paper:

@inproceedings{saporita2026pathoclass,
  title={PathoClass-BRCA: Reframing Pathology Report Generation as Guideline-Aligned Multi-Task Classification},
  author={Saporita, Alessia and Pipoli, Vittorio and Bolelli, Federico and Baraldi, Lorenzo and Acquaviva, Andrea and Ficarra, Elisa and others},
  booktitle={37th British Machine Vision Conference, BMVC 2026},
  year={2026}
}

---

References

• MI-Gen / WsiCaption — Multiple Instance Generation of Pathology Reports for Gigapixel Whole-Slide Images, MICCAI 2024.
• HistGen — Histopathology Report Generation via Local-Global Feature Encoding and Cross-modal Context Interaction, MICCAI 2024.
• BiGen — Historical Report Guided Bi-modal Concurrent Learning for Pathology Report Generation, MICCAI 2025.
• WSI-LLaVA — A Multimodal Large Language Model for Whole Slide Image, ICCV 2025.
• UNI — Towards a general-purpose foundation model for computational pathology, Nature Medicine, 2024.
• PLIP — A visual–language foundation model for pathology image analysis using medical Twitter, Nature Medicine, 2023.
• Prov-GigaPath — A whole-slide foundation model for digital pathology from real-world data, Nature, 2024.
• CLAM — Data-efficient and weakly supervised computational pathology on whole-slide images, Nature Biomedical Engineering, 2021.
• TCGA — NIH Genomic Data Commons
• ICCR — International Collaboration on Cancer Reporting, invasive breast carcinoma dataset