Speeding Up the Deep Learning Development Life Cycle for Cancer Diagnostics

Explore strategies for accelerating the deep learning development lifecycle in cancer diagnostics through this EuroPython 2021 conference talk. Learn about optimizing iteration speed, task prioritization, data processing, GPU parallelization, code quality, and continuous integration. Discover how Mindpeak creates robust deep learning models for cancer diagnostics across diverse laboratory settings. Gain insights into efficient annotations, metric definition, reproducibility, and dataset reduction techniques. Suitable for those with initial machine learning experience, this talk aims to provide actionable tips for improving model development speed and quality in medical AI applications.

Intro
Our Mission
Cancer diagnostics today
Future cancer diagnosis not for everyone?
Cancer diagnostics tomorrow
About MindPeak
Our Team and Advisors
Example: cancer cell detection
Simplicity
Training a deep learning model
Goal: Test new ideas quickly
Overview: Idea stage
Idea Generation - without data
Data-driven idea generation
Efficient Annotations
Metrics - define your target goals
Metrics - Mindpeak example
Overview: Implementation stage
Code quality-comments as code
Code quality - use einops library
On reproducibility
Implementation stage - summary
Overview: Training & Evaluation stage
PyTorch Data Parallelization
Pytorch Distributed Data Parallelization
Dataset reduction techniques
Training + evaluation stage - summary
Disappointment