Optimization of Diagnostic Radiology Imaging Models for Edge Devices

• This repository provides all the source code written for the paper Optimizing Medical Image Classification Models for Edge Devices.

Directory Information

AndroidBenchmarkingApp: Contains Android Application for benchmarking model performance on Android devices.

model_optimization: Contains source code and weights for post training optimizations and qunatization aware training. It also contains code for performance evaluation (AUC-ROC score) of models.

inference_timing: Contains source code for calculating inference time on devices (other than Android, for that use AndroidBenchmarkingApp)

NOTES:

• Change NIH_14_DATASET_PATH in all the notebooks inside model_optimization directory accordingly. This code assumes that NIH 14 datset is extracted as NIH_14 inside the model_optimization directory.

• test_images inside inference_timimg directory is radomly sampled from main dataset. Their sole purpose is to calculate timings.

• weights directory inside model_optimization contains all the weights used for this paper. One can directly use it or make their own with the help of quantize_models.ipynb notebook. The name of weight files are self explanatory.

• Weight file inside ./AndroidBenchmarkingApp/app/src/main/assets are the same as in ./model_optimization/weights and maps as:

  • QAT_int_8_v1.tflite : INT8.tflite
  • chexnet_kaggle_tflite_dynamic_quantization.tflite : dynamic_quantization.tflite
  • chexnet_kaggle_tflite_fp16.tflite : FP16.tflite

Contributing Guidelines

• Contributors are encouraged to open github issue and send their merge requests.

Contributors

• Areeba Abid (Emory University School of Medicine)
• Priyanshu Sinha (Mentor Graphics India Pvt Ltd)