Advanced Medical Computer Vision Project

Lobe Ranger: Multi-Scale Ordinal Pathology Foundation Network

A novel deep learning framework designed to process paired 20x and 40x whole-slide histopathological images from the LungHist700 dataset. Leverages a pre-trained pathology-tuned transformer backbone combined with custom bidirectional cross-scale attention fusion to simultaneously classify malignancy, carcinoma subtype (ACA vs. SCC), and ordinal tumor differentiation grade.

View Repository Model Walkthrough
90.11% Malignancy Accuracy
64.47% Subtype Accuracy
40.79% Ordinal Accuracy
1.3394 Best Validation Loss
0.9956 Malignancy ROC-AUC

Bidirectional Cross-Scale Attention Fusion

Lobe Ranger (MOPFN) integrates global tissue architecture (20x) and detailed cell cytology (40x) to mimic a pathologist's workflow of zooming in and out under the microscope.

1

Shared ViT-B Backbone

Uses a shared Vision Transformer (ViT-B/16) backbone with frozen feature extraction to keep the optimization surface focused on fusion and task heads.

2

Multi-Scale Image Pairing

Patient-wise matches 20x macro-structural and 40x micro-cytological tissue patches, capturing multi-resolution spatial features without heavy downsampling.

3

Bidirectional Attention

Enables information exchange between scales: the 20x branch queries the 40x branch for fine cytologic details, and the 40x branch queries 20x for wider contextual layout.

4

Hierarchical Ordinal Heads

Multi-task heads predict binary malignancy, subtype (masked for normal), and ordinal differentiation grade using Consistent Ordinal Regression (CORAL).

Explainable AI (XAI) & Interpretability

The repository includes both saliency maps and token-level cross-scale attention inspection so qualitative analysis can be regenerated alongside the predictive pipeline.

Malignant Case Saliency Map Focus
Malignant Case (Adenocarcinoma / Squamous Cell Carcinoma) Saliency Mapping: Highlights prominent diagnostic regions including atypical nuclear boundaries, pleomorphic nucleoli, and multi-cellular nests.
Normal Sample Saliency Map Focus
Normal Pulmonary Structure Saliency Mapping: Focuses on normal alveolar walls and architectural layouts, confirming that normal landmarks do not trigger false positive neoplastic scores.

Quantitative Evaluation Results

These values come from the latest regenerated 5-epoch local run using `outputs/improved_train_metrics.json` and `outputs/improved_eval_metrics.json` after correcting the split construction and pair explosion.

Task Metric Value
Malignancy Accuracy 90.11%
Malignancy F1 94.41%
Malignancy ROC-AUC 0.9956
Subtype Accuracy 64.47%
Subtype Macro F1 61.54%
Differentiation Accuracy 40.79%
Differentiation MAE 0.592

Caveat: this corrected split is much healthier than the earlier 142-vs-3 subtype imbalance, but it is still a 5-patient held-out set. Subtype now looks plausible, while ordinal performance remains unstable and should be cross-validated before publication.