Blurry-Consistency Segmentation Framework with Selective Stacking on Differential Interference Contrast 3D Breast Cancer Spheroid

TL;DR

This work tackles the challenge of segmenting cells in differential interference contrast (DIC) 3D breast cancer spheroids where stacking multiple z-slices introduces blur. It introduces Selective Blurry-Slice Stacking (SBS-Stacking) to curate in-focus slices, pairs it with a Dense-Stacking Consistency Mask R-CNN (DSCMask R-CNN) that enforces weak-to-strong consistency across stacked inputs, and adds Sparse-Stacking Self-Training to leverage reliable low-slice samples. The approach yields state-of-the-art segmentation performance on their DIC spheroid dataset, with notable gains in mAP metrics across several backbones and a robust ablation showing each component contributes to the improvement. Collectively, the framework provides a practical, scalable solution for accurate 3D spheroid analysis under blur, with potential extensions to tracking and deeper biological interpretation.

Abstract

The ability of three-dimensional (3D) spheroid modeling to study the invasive behavior of breast cancer cells has drawn increased attention. The deep learning-based image processing framework is very effective at speeding up the cell morphological analysis process. Out-of-focus photos taken while capturing 3D cells under several z-slices, however, could negatively impact the deep learning model. In this work, we created a new algorithm to handle blurry images while preserving the stacked image quality. Furthermore, we proposed a unique training architecture that leverages consistency training to help reduce the bias of the model when dense-slice stacking is applied. Additionally, the model's stability is increased under the sparse-slice stacking effect by utilizing the self-training approach. The new blurring stacking technique and training flow are combined with the suggested architecture and self-training mechanism to provide an innovative yet easy-to-use framework. Our methods produced noteworthy experimental outcomes in terms of both quantitative and qualitative aspects.

Figures (3)

• Figure 1: Our proposed pipeline for stacking and training 3D cell images. The 3D microscopy images are, firstly, stacked before running through the training stage of the proposed architecture DSCMask R-CNN. Finally, a self-training framework is employed to select reliable samples for the continuous training epochs.
• Figure 2: Proposed pipeline in detail with two main parts. 1) SBS-Stacking takes the original dataset to create partially stacked images and full focus-stacked images. 2) DSCMask R-CNN architecture that utilizes Mask R-CNN and mechanism of consistency training by treating SBS-Stacking images as a weak-augmented set and Focus-Stack images as a strong-augmented set. 3) Sparse-stacking consistency with a Self-training scheme will filter and select reliable images for the re-training stage by using a teacher model.
• Figure 3: Visualization of a) Out-of-focus slice ($Z$-13); b) Partially in-focus slice ($Z$-6); c) and d) In-focus detection map of image b) and its overlay image; e) and f): Output images of SBS-Stacking; g) and h): Output images of normal Focus-Stack.