Thoughts on the New “Deep Learning with PyTorch” (2nd Edition)

I’ve been eagerly awaiting the release of the second edition of Deep Learning with PyTorch, and finally grabbed my copy as soon as it came out on March 10th.

Given my current work building models for early disease recognition, the very first thing I did was flip straight to Chapter 11 to see how the authors handle medical imaging. I quickly noticed that they advocate for the slicing approach—extracting individual 2D cross-sections from a full scan.

Personally, I tend to prefer resizing the full 3D images directly. Slicing inherently forces the data from 3D down to 2D, which means losing some of the depth spatial context. By resizing the entire 3D volume, you preserve the vital relationships between structures across all planes (axial, sagittal, and coronal), which I’ve found crucial for detecting volumetric anomalies.

That being said, I’m keeping an open mind. Every approach has its trade-offs, and I’m really looking forward to diving deeper into their methodology and learning some new techniques from the book!