Indifference subspace of deep features for lung nodule classification from CT images

Abstract

Deep learning (DL) has made substantial contributions to automated diagnoses in biomedical imaging, with various architectures extensively used for critical classifications such as lung nodule detection from CT scans. Despite satisfactory results from basic DL implementations, understanding DL’s inner mechanisms and parameter evolution remains understudied. DL layers typically favor nodes with larger activation values, facilitating a softmax-type decision post-training. This aligns with various alternative final-layer replacements like support vector machines (SVM), random forest, naive Bayes, and k-nearest neighbor (k-NN). However, replacing the decision layer with a classifier that operates in the so-called indifference subspace, like the common vector approach (CVA), may disrupt the standard paradigm, as it requires commonality in feature node magnitudes rather than large feature values. This study investigates the feasibility of adapting standard DL architectures to generate feature nodes with common magnitudes conducive to CVA fine-tuning. Surprisingly, we find that DL networks, even without explicit design for this purpose, can achieve remarkable classification accuracies through CVA, effectively on par with state-of-the-art results. The intriguing high classification accuracy is examined through the relationship between “indifference subspace” and “node value,” scrutinized via an expansive suite of DL architectures, with and without ImageNet pre-training. Although the aim of the study is limited to the possibility of subspace alignment in the feature layers of convolutional neural networks (CNNs), the results demonstrate that CVA fine-tuning not only challenges the prevailing paradigms within DL classifications but also unveils a novel pathway for possibly enhancing classification performance in biomedical imaging, particularly for lung nodule detection.