leonvol / boneage

Experiments and submission code of project 'BoneAge' for competition BWKI

This project is the world's first AI system to determine a person's age by analyzing 3D low-dose thorax CT images of the clavicle. It has higher accuracy and a wider age detection range than more traditional hand bone age assessment and is much faster than estimates of trained radiologists.

→ Invitation to 2020's nationwide final, placed TOP 5

Code structure overview

module name	function
batch_loader	fast, parallelized loading, processing, augmenting and caching of CT images
train_framework	framework to train and compare the performance of different net structures
vgg16_3d	implementation of a 3D VGG16 Net
vgg16_attention_pretrained	pretrained 3D VGG16 Net with attention
alexnet_3d	implementation of a 3D Alexnet
convert_crop	automatically crop and convert DICOM data with segmentation point
preprocessing	helper functions for preprocessing
util	general helper functions
clr_callback	cyclic learning rate callback for keras
predict	prediction of not yet segmented CT images

Installation

Installation of all needed dependencies by running

pip install -r requirements.txt

Results

The best models can be downloaded from Google Drive

	neural net structure	learning rate	Test-Set MAE in months
1	3D VGG16, BN, 3 Dense*	CLR [0.01, 0.001]	23.14
2	3D AlexNet, 4 Conv Layers, BN, 3 Dense	CLR [0.01, 0.001]	23.76
3	3D VGG16, BN, GlobalMaxPooling3D*	CLR [0.01, 0.001]	25.60
4	VGG16 Attention**, ersten 3 Layer trainierbar, BN, 3 Dense	CLR [0.1, 0.01]	30.16
5	VGG16 Attention**, GlobalMaxPooling	CLR [0.1, 0.01]	32.43
...

*modified, without pooling after the 4th block to allow for convolutions in the 5th block

**pretrained on RSNA Bone Age from kaggle

Acknowledgment

Thanks to LMU for the dataset