BlobCUT: A Contrastive Learning Method to Support Small Blob Detection in Medical Imaging

Abstract

Medical imaging-based biomarkers derived from small objects (e.g., cell nuclei) play a crucial role in medical applications. However, detecting and segmenting small objects (a.k.a. blobs) remains a challenging task. In this research, we propose a novel 3D small blob detector called BlobCUT. BlobCUT is an unpaired image-to-image (I2I) translation model that falls under the Contrastive Unpaired Translation paradigm. It employs a blob synthesis module to generate synthetic 3D blobs with corresponding masks. This is incorporated into the iterative model training as the ground truth. The I2I translation process is designed with two constraints: (1) a convexity consistency constraint that relies on Hessian analysis to preserve the geometric properties and (2) an intensity distribution consistency constraint based on Kullback-Leibler divergence to preserve the intensity distribution of blobs. BlobCUT learns the inherent noise distribution from the target noisy blob images and performs image translation from the noisy domain to the clean domain, effectively functioning as a denoising process to support blob identification. To validate the performance of BlobCUT, we evaluate it on a 3D simulated dataset of blobs and a 3D MRI dataset of mouse kidneys. We conduct a comparative analysis involving six state-of-the-art methods. Our findings reveal that BlobCUT exhibits superior performance and training efficiency, utilizing only 56.6% of the training time required by the state-of-the-art BlobDetGAN. This underscores the effectiveness of BlobCUT in accurately segmenting small blobs while achieving notable gains in training efficiency.

Keywords: Hessian analysis; blob detection; contrastive learning; glomeruli segmentation; imaging biomarker.

Figure A1

Spherical coordinate systems.

Figure 1

Steps to identify blobs by using BlobCUT.

Figure 2

The training process of our proposed BlobCUT.

Figure 3

Illustration of blob identification through joint constraint operation. (a) Original noisy blobs image. (b) Clean blob image of blobs without noise. (c) Ground truth of blob centers. (d) Hessian convexity mask. (e) Blob mask from networks. (f) Final blob identification mask.

Figure 4

Illustration of the training datasets used by different experiments. For Exp. I, synthetic noisy blob images were used as the source domain images; for Exp. II, real kidney MR images were used as the source domain images. For both experiments, synthetic clean blob images were used as target domain images to encourage the model to have denoising capability.

Figure 5

Illustration of training input images of BlobCUT (a) Synthesized 3D blobs image from domain clean 3D blobs. (b) Blob mask of (a). (c) Synthesized 3D noisy blobs image from domain noisy 3D blobs. (d) Synthesized 3D mouse kidney image patch from domain noisy 3D blobs. (e) Real 3D mice kidney image patch from domain T.

Figure 6

Illustration of the learning curve. (a) Loss curve comparison between training and validating of BlobCUT. (b) Testing F-score comparison between BlobCUT and BlobDetGAN.

Figure 7

Comparison of glomerular segmentation results from 3D MR images of mouse kidneys using BlobCUT, BlobDetGAN, UH-DOG, BTCAS, UVCGAN and EGSDE. Identified glomeruli are marked in red. Three slices are illustrated: kidney #429 slice 111, #466 slice 96 and #469 slice 81.

Figure 8

Denoising results of noisy synthetic blobs image using U-Net, 3D CUT, BlobCUT and compared with ground truth. (a) Original noisy blobs image. (b) Ground truth. (c) Denoised result of U-Net. (d) Denoised result of 3D CUT. (e) Denoised result of BlobCUT.