CNNs vs Intensity-Based Features

You’re working as a biological data scientist at Columbia University on a research project that aims to automatically detect early-stage lung abnormalities from chest X-ray images. A senior researcher on the team suggests building a model based on handcrafted intensity-based markers, such as average pixel brightness in specific lung regions and simple texture statistics.

Another group proposes using a convolutional neural network (CNN) to learn features directly from the raw images, arguing it could capture more complex spatial patterns.

Both approaches have shown some promise in small pilot experiments, but you need to recommend a direction for the full study, where model interpretability, generalization across hospitals, and robustness to imaging differences are all important.

How would you evaluate these two modeling strategies? What factors would guide your investigation, and what evidence would you look for before deciding whether a CNN-based approach is justified over simpler intensity-based features?