Multi-headed U-Net: an automated nuclei segmentation technique using Tikhonov filter-based unsharp masking

An automated nuclei segmentation is the key technique for understanding and analyzing cellular properties, which are helpful for disease diagnosis and support computer-aided digital pathology. However, this task is challenging because of the variability in nuclei size and morphology, which results in blurry boundaries and overlapping nuclei. To address such issues, a multi-headed U-Net convolutional neural network (CNN) architecture has been proposed. This architecture has multiple heads to extract multi-resolution features of the source image by using different kernel sizes of the filters. The source images are pre-processed using an unsharp masking approach based on the Tikhonov filter. The Tikhonov filter decomposes the input image into low-frequency and high-frequency band images. The unsharp masking method improves the high-frequency information of the input image by primarily enhancing features such as boundaries, contours, and fine details. We have incorporated intersection over union (IOU) and F1Score as measures along with accuracy for our proposed objective functions. The proposed objective functions are tried to be maximized by the optimization algorithm, and the higher value of the metrics indicates better segmentation performance in the spatial domain during the testing phase. The proposed method attained IOU(JI), Accuracy, Precision, and F1Score values as 0.8299, 0.9642, 0.8918, and 0.9070, respectively. The quantitative and qualitative experimental outcomes indicate that our proposed technique outperforms the state-of-the-art techniques.

自动化细胞核分割是理解与分析细胞特性的核心技术，可为疾病诊断提供助力，并支撑计算机辅助数字病理（computer-aided digital pathology）领域的发展。然而，该任务颇具挑战性：细胞核的尺寸与形态存在显著差异，进而导致边界模糊与细胞核相互重叠的问题。针对上述问题，本研究提出了一种多头U-Net卷积神经网络（CNN）架构，该架构通过采用不同尺寸的卷积核，可提取源图像的多分辨率特征。研究采用基于蒂霍诺夫（Tikhonov）滤波器的非锐化掩蔽（unsharp masking）方法对源图像进行预处理，蒂霍诺夫滤波器可将输入图像分解为低频波段图像与高频波段图像；非锐化掩蔽方法主要通过增强边界、轮廓与细微细节等特征，提升输入图像的高频信息。本研究所提出的目标函数，同时纳入交并比（IOU）、F1分数（F1Score）以及准确率作为评估指标，并通过优化算法对所提目标函数进行最大化求解。在测试阶段，各项评估指标的数值越高，则代表空域下的分割性能越优异。本方法的交并比（IOU，亦称JI）、准确率、精确率以及F1分数分别为0.8299、0.9642、0.8918与0.9070。定量与定性实验结果表明，本研究所提方法的性能优于当前前沿技术。