Research on interpretable CNN classifiers, involves comparing semantic segmentation masks with heat maps designed as visual explanations. A robust explanation accurately identifies or approximates the segmentation of an object. Our focus is on CNN classifiers with enhanced explainability, particularly in the middle layers. To explore this, we propose testing an encoder, trimmed to a medium layer, within a Fully Convolutional Network (FCN). Semantic segmentation is a pivotal task in computer vision preceding object recognition, and demands efficiency to optimize performance, energy consumption, and hardware costs. While various lightweight FCN proposals exist for distinct semantic segmentation tasks, their designs often introduce additional complexity compared to the more basic FCN design we advocate. Our goal is to see how well a minimal FCN works in a simple semantic segmentation task using medical images and how its accuracy changes when the training dataset is shrunk. The study involves characterizing and comparing our minimal FCN against other lightweight deep segmentation models and analyzing accuracy curves concerning the quantity of training data. Utilizing chest CT imaging, we focus on segmenting the lungs. We highlight the importance of data consumption and model size as decisive factors in selecting an architecture, especially when differences in predictive accuracy are marginal. Characterizing deep architectures based on their data requirements, allows for a thorough comparison fostering a deeper understanding of their suitability for specific applications.
|