AMLNet

This code implements AMLNet from the manuscript "AMLnet, A deep-learning pipeline for the differential diagnosis of acute myeloid leukemia from bone marrow smears".

Prerequisites

Any operating system on which you can run GPU-accelerated PyTorch. Python >=3.8. For packages see requirements.txt.

Training

Directory structure for training and test data:

DATA
└── Train
│   ├──M1
│   ├──M2a
│   ├──...
└── Test
│   ├──M1
│   ├──M2a
│   ├──...

For 5-fold cross-validation train

DATA
└── All
│   ├──M1
│   ├──M2a
│   ├──...

Code in ./train

1.train.py: it uses early stop, label smoothing, various data augmentation methods, cosine annealing with warm restart and so on for training.
2.train_kfold.py: based on the above, including 5-fold cross-validation for training.

Configuring hyperparameters via the ./experiments directory

./train : compare different models.
./kfold : implement 5-fold cross-validation, add DROPOUT under the MODEL directory.

Nvidia TITAN RTX with 24GB of memory for GPU is used to accelerate training.

Test

For test at the image-level

DATA
└── Dual_center_test
│   ├──M1
│   ├──M2a
│   ├──...

For test at the patient-level

DATA
└── Dual_center_test
│   ├──M1
	│   ├──Patient1
	│   ├──Patient2
	│   ├──...
│   ├──M2a
	│   ├──Patient1
	│   ├──Patient2
	│   ├──...
	...

Code in ./test

1.test.py: evaluate model performance at the image level.
2.test_patient.py: evaluate model performance at the patient level.

Publicly available software tool

Local test, open to acess soon

Available on mobile phone

Available on website

bigbins/AMLNet-model