Segmentation & classification of embryo videos in in-vitro fertilization
summary
Pipeline segments embryos with U-Net segmentation trained on OpenCV-derived masks and achieves near‑perfect cropping, then learns image embeddings with a stacked denoising autoencoder (preferred to a blurrier VAE) and feeds downsampled sequences into an LSTM 500–1,000 units classifier producing a probability over liveborn/miscarriage/no‑implantation (classes later merged). Data is a few thousand labeled videos plus many unlabeled; segmentation works well, classification shows usable accuracy but needs much more diverse data and end‑to‑end fine‑tuning with patient records to improve clinical reliability.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 5:21Deep Learning
- 6:35Architecture
- 8:48Extraction
- 11:01Unit
- 15:45Unsupervised Feature
- 16:57Variational Autoencoder
- 19:19Noising Out Encoder
- 21:52Stacked Noising Out Encoder
- 23:01Video Classification
- 25:44Feature Extraction
- 27:43Results
- 29:15Future work
related talks
45mTech Tutorial: Image segmentation
55mDeep Learning in Medical Image Diagnostics
36mDeep Learning - Segmentation, Identification and Classification
34mWhy AI in the Medical Field Both Sucks and Rocks...
22mSemantic segmentation of tennis match videos
49mBuilding AI That Learns and Adapts: A Case Study in MRI Diagnostics
43mBuilding AI That Learns and Adapts: A Case Study in MRI Diagnostics
11mInterview