danielaxv3/lungcv-xray-multilabel-v32

LungCV v3.2 — Multitask + Multi-label (production)

Multi-label chest X-ray finding classifier with an auxiliary lung-segmentation head: v2's proven segmentation-aided training recipe + v3's independent sigmoid output design + OOD rejection. Each finding (NORMAL / PNEUMONIA / COVID / TB) is scored independently; films where nothing crosses the reporting threshold are flagged uncertain; films far from the training distribution are flagged out-of-distribution.

Research prototype only — not a medical diagnosis.

Head-to-head vs v2 (864 held-out tiles, identical inputs, same patient-wise split)

Metric	v2 multitask	v3.2
Argmax accuracy	88.4%	89.8%
Macro-F1	0.892	0.903
NORMAL	68.0%	71.6%
COVID	96.0%	97.6%
PNEUMONIA	99.6%	99.6%
TB	92.1%	91.2%
Unfamiliar films (5 CC0)	1/5 (20%)	3/5 (60%)
Uncertain-flag rate (familiar)	— (always answers)	6.4%

Own-validation (1,800 images): 89.4% accuracy, macro-F1 0.894.

Architecture

• Backbone: EfficientNetB0 (ImageNet), frozen head training then two-stage fine-tune (top 40 encoder layers unfrozen, BatchNorm frozen, Adam 1e-5)
• Heads: GAP → Dropout(0.3) → Dense(4, sigmoid) (binary cross-entropy) + U-Net-style segmentation decoder (weight 0.3, ground-truth + U-Net pseudo-masks)
• Training data: 12,000 balanced tiles (3,000/finding) from 7 public datasets, patient-wise 70/15/15 split (seed 42)
• OOD: per-class centroids in the 1,280-d GAP feature space; threshold 9.85 (95th percentile of train distances)

Preprocessing contract

1. Grayscale
2. Resize 224×224 (no crop, no CLAHE)
3. Replicate to 3 channels, raw [0,255] (EfficientNetB0 built-in rescaling)

Files

• best_v32.keras — the model (77 MB, classifier + segmentation head)
• metrics.json, preprocessing_config.json (includes ood_distance_threshold), class_names.json, ood_centroids.json, confusion_matrix.png

Model history

Version	Design	Familiar acc (shared cohort)	Unfamiliar
v1	EfficientNetB0 softmax	—	—
v2	multitask softmax	88.4%	20%
v3	DenseNet121 + CLAHE, sigmoid, no seg head	79.8%	60%
v3.2	multitask + sigmoid + OOD	89.8%	60%

The v3 experiment isolated the lessons that built v3.2: CLAHE hurt NORMAL recall (kept out), the seg-head learning signal matters (brought back), the sigmoid + OOD design is the differentiator (kept).