NCSAS 2026 — AI-Based Leukemia Detection: Model Extension

Authors: Ayush Patel, Yuval Dinodia, Rohin Patel

Fine-tuning DinoBloom-G (ViT-G/14, 1.1B parameters) for leukemia subtype classification, with cross-institutional generalization experiments and knowledge retention analysis.

---

Overview

Standard CNNs trained on a single institution's data suffer severe accuracy drops when tested across institutions (~23 pp in our experiments). This project fine-tunes DinoBloom-G — a foundation model pretrained on 380K hematology images — on multi-domain leukemia data to achieve robust performance.

Model	Accuracy
Linear probe (frozen DinoBloom-G)	89.41%
Fine-tuned DinoBloom-G (ours, 75 epochs)	98.86%
EfficientNet-B0 in-distribution	95.12%
EfficientNet-B0 cross-institution	71.85%

---

Scripts

train_efficientnet_b0.py

Trains an EfficientNet-B0 classifier on leukemia cell images using an ImageNet-pretrained backbone with a trainable classification head. Supports partial unfreezing of the last N feature blocks.

Key flags:

Flag	Default	Description
--epochs	30	Training epochs
--lr	1e-4	Learning rate
--unfreeze-blocks	2	Feature blocks to unfreeze (0–9)
--batch-size	32	Batch size

---

single_institution_cnn.py

Establishes the cross-institutional generalization baseline. Trains an EfficientNet-B0 head on C-NMC (India) and evaluates on Taleqani Hospital (Iran).

Split	Accuracy
In-distribution (C-NMC)	95.12%
Cross-institution (Taleqani)	71.85%
Drop	−23.27 pp

Outputs to single_institution_results.txt.

---

linear_probe.py

Evaluates the general WBC recognition capability of the original DinoBloom-G backbone by training a frozen linear classifier (L-BFGS) on extracted 1536-dim features. Serves as the pre-fine-tuning baseline.

python linear_probe.py --backbone DinoBloom-G.pth --data-dir <path>

Outputs to linear_probe_results.txt.

---

compare_retention.py

Measures how much general hematological knowledge is retained after fine-tuning. Runs a 5-NN classifier on features from both the original DinoBloom-G and the fine-tuned backbone across 21+ WBC cell types.

python compare_retention.py --original DinoBloom-G.pth --finetuned checkpoint_latest.pth

Outputs to retention_<timestamp>.txt.

---

eval_val.py

Evaluates a trained checkpoint on the validation split, reporting per-class and macro precision, recall, and F1.

python eval_val.py --checkpoint bloom_leukemia.pth --workers 0

---

test_model.py

Evaluates a trained checkpoint at both the image level and the patient level (majority vote). Requires patient_metadata.csv mapping patient IDs to diagnoses.

python test_model.py --checkpoint bloom_leukemia.pth --data-dir <path>

Outputs to test_results_<timestamp>.txt.

---

plot_training.py

Reads training_metrics.csv and generates 4 separate training curve plots: loss, accuracy, train-val gap, and learning rate schedule.

python plot_training.py

---

Model Architecture

DinoBloom-G (ViT-G/14, pretrained on 380K hematology images)
  Transformer blocks 0 to N-3   [FROZEN]
  Transformer blocks N-2 to N   [TRAINABLE]
        |
        Classification Head
              LayerNorm(1536)
              Linear(1536 -> 512)
              GELU + Dropout(0.2)
              Linear(512 -> num_classes)

Training setup: AdamW, CosineAnnealingLR, label smoothing 0.1, bfloat16 AMP, 75 epochs

---

Requirements

• Python 3.10+, PyTorch 2.x with CUDA
• torchvision, Pillow, matplotlib
• DinoBloom-G.pth (original pretrained weights)
• dinov2/ package in repo root