Leukemia Classification with Double Transfer Learning

Abstract

This project implements an improvised double transfer learning pipeline for Acute Lymphoblastic Leukemia (ALL) classification using microscopic Peripheral Blood Smear (PBS) images.
We leverage ResNet-50, EfficientNet-B3, and InceptionV3 backbones with ImageNet pretraining.
Two-stage fine-tuning is applied:

• Phase-1: Binary classification (Normal vs Leukemia) on Dataset-1 (ISBI ALL Challenge).
• Phase-2: Multi-class classification (Benign vs Early Pre-B, Pre-B, Pro-B ALL) on Dataset-2 (Kaggle ALL PBS).

This successive transfer improves performance on small medical datasets by progressively adapting the models to domain-specific features.

---

Dataset Details

Dataset-1 (ALL Challenge — TCIA)

• Images: 15,135
• Patients: 118
• Classes: Normal Cells, Leukemia Blasts
• Source: The Cancer Imaging Archive (TCIA)
• Citation: Gupta, A. & Gupta, R. (2019). ALL Challenge dataset of ISBI 2019. DOI: 10.7937/tcia.2019.dc64i46r

Dataset-2 (ALL PBS — Kaggle)

• Images: 6,512 (segmented PBS images)
• Patients: 89
• Classes: Benign (Hematogones), Malignant (Early Pre-B, Pre-B, Pro-B)
• Source: Kaggle
• Citation: Mehrad Aria et al. “Acute Lymphoblastic Leukemia (ALL) image dataset.” Kaggle, 2021. DOI: 10.34740/KAGGLE/DSV/2175623

---

Methodology

Data Augmentation

• Random rotations, flips, crops, and zooms
• Color jitter (brightness/contrast/saturation)
• HSV segmentation preprocessing for PBS images

Double Transfer Learning

1. Primary Transfer Phase
   Train ResNet50, EfficientNetB3, InceptionV3 (ImageNet → Dataset-1).
   Focus: learn leukemia vs normal morphology.

2. Secondary Transfer Phase
   Fine-tune the same backbones on Dataset-2 (specialized PBS ALL subtypes).
   Added custom CNN layers for PBS-specific features.

---

Implementation

• Framework: PyTorch
• Pretrained Models: ResNet-50, EfficientNet-B3, InceptionV3
• Optimizer: AdamW / Adamax
• Loss: CrossEntropy
• Scheduler: CosineAnnealingLR
• Mixed Precision: Enabled (AMP)

---

Results (Reported)

Model	Phase-2 Accuracy	Eval Acc	Notes
EfficientNetB3	95.98%	90.23%	Best performing
ResNet-50	83.76%	95.50%	Strong generalization
Inception-V3	44.1% (10 epochs)	86.25%	Underfit with short training

---

Novel Extension (Future Work)

• Integrating Fuzzy Inference System (Mamdani) with features from double transfer learning models.
• Aim: Improve robustness on unseen data with fuzzy clustering.

---

How to Run (This Repo)

Phase-1 Training

python -m src.train_phase1 --train_dir /path/to/ds1/train --val_dir /path/to/ds1/val --arch resnet50 --epochs 10 --out_dir checkpoints

Phase-2 Fine-Tuning

python -m src.finetune_phase2 --train_dir /path/to/ds2/train --val_dir /path/to/ds2/val --phase1_ckpt checkpoints/phase1_resnet50_best.pt --epochs 10 --out_dir checkpoints

Evaluation

python -m src.evaluate --test_dir /path/to/ds2/test --ckpt checkpoints/phase2_resnet50_best.pt

Predict Single Image

python -m src.predict_one --image path/to/sample.jpg --ckpt checkpoints/phase2_resnet50_best.pt

---

References

• Gupta, A., & Gupta, R. (2019). ALL Challenge dataset of ISBI 2019 [Data set]. TCIA. DOI: 10.7937/tcia.2019.dc64i46r
• Mehrad Aria et al. (2021). Acute Lymphoblastic Leukemia (ALL) image dataset. Kaggle. DOI: 10.34740/KAGGLE/DSV/2175623
• Ghaderzadeh, M., Aria, M., Hosseini, A, Asadi, F, Bashash, D, Abolghasemi, H. (2022). A fast and efficient CNN model for B-ALL diagnosis. Int J Intell Syst, 37: 5113–5133. DOI: 10.1002/int.22753