Recurrent Multi-Scale U-Nets for Improved Medical Image Segmentation: A Comprehensive Investigation

Medical image segmentation is a critical task in healthcare, with numerous challenges stemming from the variability of medical images, the complexity of the structures being segmented, and the need for precise accuracy. Segmentation problems often suffer from noisy data, irregular shapes, and overlapping structures, which make it difficult to achieve both high accuracy and topological consistency. Moreover, the computational cost of processing large, high-dimensional datasets poses further limitations. This thesis addresses these issues by proposing novel approaches that enhance the performance and robustness of segmentation models, focusing on improving segmentation accuracy, maintaining topological consistency, and optimizing computational efficiency across diverse datasets.

In this research, we solve two different medical image segmentation problems. A contributions chapter has been dedicated to each. First, we achieve leading performance according to Sensitivity (SEN), Jaccard Index (JC), and F1 scores on Phase Contrast Microcopy datasets within the Cell Tracking Challenge (CTC). We select the 2D+Time datasets for our research for which both Gold and Silver Truths were available, amounting to 8 datasets. Cell Microscopy Segmentation is a field of research with practical research and diagnostic merit. This can help to track the movement of target cell types, diagnose neurodegenerative diseases, identify malignant cell growths, or detect abnormal blood cells (sickle cell anemia). We achieve these results using our novel approach of Weighted Spatial DICE (WSD) Loss to train our new architecture leveraging recurrent connections: Bidirectional Squeeze Excitation Modified Attentional Recurrent Multi-Scale-UDet (BiSE-MARM-UDet).

For our second contributions chapter, we address the Retina Blood Vessel Segmentation problem. This problem is of practical value to solve as accurate identification of the vascular tree network can help identify diseases such as Retinopathy, Glaucoma, Retinal Detachment, and Coroidal Tumours. Blood vessels are contiguous structures however, and diseases identification need to connective topology of the vascular tree to be correct. Thus we specialize BiSE-MARM-UDet for topologically accurate segmentation of fine-scaled structures, introducing Bidirectional-Multi-Scale-Attentional-Recurrent-UDet (BiMAR-UDet). Additionally, WSD Loss is improved using electric field divergence to yield the Topology-Aware Loss Function Weighted-Spatial-Electrical-Divergence-DICE (WSEDD) Loss. We train and evaluate our methods using the fundus image datasets: DRIVE, STARE, IOSTAR, and CHASEDB1. This model achieves leading performance by not only SEN, JC, and F1 scores, but also Betti Error (BE) as a topological metric. BE measures the sum in error in the Betti numbers β0 and β1.

In both contributions chapters, we have conducted extensive ablation studies pertaining to both our Architectural and Loss contributions. We provide extensive evidence that our methods are effective, and cutting edge at solving both segmentation problems.