Actin-Villin Fiber Segmentation: Use ML and Statistical Tests For Missing Fiber Models?

A Case Study from Cryo-EM tomography

Abstract

This paper addresses the longstanding challenge of distinguishing relevant regions from background in images without prior knowledge, leveraging class statistical separability alongside three machine learning methods. The approach is demonstrated on a Cryo-EM molecular tomography scan of actin–villin fibers using three classification techniques: Logistic Regression for binary outcomes, K-Means for basic clustering, and hierarchical clustering for flexible merging. The methods are evaluated for consistency across varying classification parameters and validated using statistical tests. The analyzed image exhibits moderate texture with elongated segments that intermix but can be further partitioned into multiple clusters. Segmentation results are internally validated by varying execution parameters, demonstrating the robustness of the approach. These methods highlight applications in unsupervised segmentation and can be extended to segment videos or image sequences with minimal input.