Bone X-Ray Deep Learning Competition

Path on server: /DATA4_DB3/data/sqpeng/Projects/MURA

start visdom: python -m visdom.server

LOGs

🌀 2018-06-11 13:02

Accuracy=66.31%

🌀 2018-06-12 08:09

Accuracy=71.60%

🌀 2018-06-13 01:45

Accuracy=83.95%

🌀 2018-07-10

Average:

model	ELBOW	FINGER	FOREARM	HAND	HUMERUS	SHOULDER	WRIST	AVERAGE
Baseline	0.710	0.389	0.737	0.851	0.600	0.729	0.931	0.705
DenseNet169	0.746	0.582	0.711	0.540	0.778	0.576	0.701	0.662
ResNet152	0.733	0.569	0.650	0.480	0.793	0.576	0.710	0.644
VGG16	0.632	0.592	0.650	0.495	0.689	0.566	0.742	0.624
Ensemble(3)	0.760	0.569	0.681	0.480	0.807	0.586	0.729	0.659
Ensemble(2)	0.760	0.546	0.665	0.525	0.793	0.586	0.719	0.656

Min:

model	ELBOW	FINGER	FOREARM	HAND	HUMERUS	SHOULDER	WRIST	AVERAGE
Baseline	0.710	0.389	0.737	0.851	0.600	0.729	0.931	0.705
DenseNet169	0.764	0.608	0.683	0.545	0.763	0.579	0.770	0.673
ResNet152	0.735	0.665	0.696	0.533	0.733	0.557	0.714	0.662
VGG16	0.614	0.561	0.682	0.507	0.674	0.518	0.712	0.610
Ensemble(3)	0.734	0.594	0.742	0.555	0.719	0.578	0.750	0.667
Ensemble(2)	0.774	0.665	0.697	0.555	0.748	0.588	0.733	0.680

1. 数据预处理。
2. 考虑 Collaborative Learning 。
3. 考虑 各个部位共享一部分网络，然后单独训练。

🌀 2018-07-20 16:15

完成 MultiBranchDenseNet169，对于每个部位，分别训练最后一个DenseBlock和Classifier，但是效果不好，可能还存在问题。

Dataset

Path: /DATA4_DB3/data/public/MURA-v1.1

Description:

• 7 folders: XR_ELBOW, XR_FINGER, XR_FOREARM, XR_HAND, XR_HUMERUS, XR_SHOULDER, XR_WRIST

• Training data: 36808 images, 13457 cases

• Validation data: 3197 images, 1199 cases

• Training Data Distribution

body parts	positive	negative	total
ELBOW	2006	2925	4931
FINGER	1968	3138	5106
FOREARM	661	1164	1825
HAND	1484	4059	5543
HUMERUS	599	673	1272
SHOULDER	4168	4211	8379
WRIST	3987	5765	9752
ALL	14873	21935	36808

• Test Data Distribution

body parts	positive	negative	total
ELBOW	230	235	465
FINGER	247	214	461
FOREARM	151	150	301
HAND	189	271	460
HUMERUS	140	148	288
SHOULDER	278	285	563
WRIST	295	364	659
ALL	1530	1667	3197

Images in training set are grey-scale images, i.e., there is only one channel.

The sizes of these images are various. Minimum width is 89, and minimum height is 132. Maximum width is 512, and maximum height is 512.

References

• PyTorch Book