-
Notifications
You must be signed in to change notification settings - Fork 1
Home
Ethan Weinberger edited this page May 10, 2018
·
24 revisions
-
https://www.biorxiv.org/content/biorxiv/early/2017/05/12/064279.full.pdf
- Goal: Use deep learning to predict SPOP gene mutations in prostate cancer cells
- Used an ensemble of ResNet-50 models modified with an additional dropout + fully connected layer.
- Networks were all pretrained on ImageNet data.
- Each network in the ensemble was trained on a small draw of slide data (had 157 patient slides in total).
- Weak networks (weak since they were trained on little data) were combined to form a strong learner ensemble
- Caveat: Paper doesn't provide indication of how well/poorly trained pathologists perform on this data
-
http://journals.plos.org/plosone/article/file?id=10.1371/journal.pone.0192726&type=printable
- Goal: Identify patients with clinical heart failure using H&E tissue
- Dataset: 209 patients
- 94 with heart failure
- 115 w/o failure
- Compared CNN to traditional feature-engineering model
- Pathologists classify correctly with 75% accuracy - CNNs beat this with 20% increase on sensitivity and specificity
- Architecture was modified from that of Janowczyk and Madabhushi (these guys do lots of DL pathology work)
-
https://arxiv.org/pdf/1802.09424.pdf
- Investigated the transfer learning performance of InceptionV3 and ResNet-50 with H&E images