A Suitability Assessment Framework for Medical Cell Images in Chromosome Analysis

The process of chromosome karyotype analysis is a highly time-consuming and error-prone task heavily relying on the experience of the cytogeneticists and influenced by factors such as fatigue and decrease of attention. Many efforts have dedicated to automatic chromosome karyotype analysis using various computer vision techniques based on geometric morphology and deep learning. However, few of them have paid attention to selections of high-suitability medical cell images for chromosome karyotype analysis. High-suitability cell images not only can significantly decrease the difficulty of manual chromosome karyotype analysis, but also can boost the analysis performance of automatic chromosome karyotype analysis algorithms. This paper proposes a suitability assessment framework for evaluating the suitabilities of cell images to address the issue of selecting high-suitability medical cell images for the inputs of chromosome karyotype analysis. The quantitative experimental results show that using the proposed suitability assessment framework to select suitable inputs can significantly boost chromosome segmentation performance by 5.06 percentage points of mAP, 2.4 percentage points of $A{P}^{50}$, and 3.58 percentage points of $A{P}^{75}$. The qualitative experiments with a group of cell images show that the corresponding suitability results evaluated by the proposed framework are highly in accordance with results evaluated by the experienced analysts, demonstrating the effectiveness of the proposed method to address the selection issue of suitable medical cell images.