PanNuke Dataset Extension, Insights and Baselines

TL;DR

This work introduces PanNuke, a large, diverse, and clinically QCed nucleus dataset spanning 19 tissues and 5 nucleus classes to better reflect real-world whole-slide images. It describes a semi-automatic ground-truth generation pipeline using NuClick for segmentation masks, yielding 189,744 nuclei across 481 visual fields from 20k+ WSIs at 40×, and provides a detailed taxonomy and tissue-wise statistics. The authors benchmark several nucleus segmentation and classification models, with HoVer-Net generally delivering the strongest performance (measured by multi-class PQ and related metrics) and demonstrate partial generalization to unseen tissues such as brain. They also discuss limitations such as artifacts and hidden stratification, and offer schemas and recommendations to guide robust, cross-tissue DL development in computational pathology.

Abstract

The emerging area of computational pathology (CPath) is ripe ground for the application of deep learning (DL) methods to healthcare due to the sheer volume of raw pixel data in whole-slide images (WSIs) of cancerous tissue slides. However, it is imperative for the DL algorithms relying on nuclei-level details to be able to cope with data from `the clinical wild', which tends to be quite challenging. We study, and extend recently released PanNuke dataset consisting of ~200,000 nuclei categorized into 5 clinically important classes for the challenging tasks of segmenting and classifying nuclei in WSIs. Previous pan-cancer datasets consisted of only up to 9 different tissues and up to 21,000 unlabeled nuclei and just over 24,000 labeled nuclei with segmentation masks. PanNuke consists of 19 different tissue types that have been semi-automatically annotated and quality controlled by clinical pathologists, leading to a dataset with statistics similar to the clinical wild and with minimal selection bias. We study the performance of segmentation and classification models when applied to the proposed dataset and demonstrate the application of models trained on PanNuke to whole-slide images. We provide comprehensive statistics about the dataset and outline recommendations and research directions to address the limitations of existing DL tools when applied to real-world CPath applications.

Figures (11)

• Figure 1: Illustration of PanNuke label generation and verification.
• Figure 2: 1st column: A selection of visual fields from the Kumar dataset et al.kumar_dataset_2017. 2nd-3rd columns: A selection of visual fields in PanNuke with output of a detector trained on kumar_dataset_2017 overlaid on the images. False positives (shown as red dots, as opposed to true positives as yellow dots) are clearly visible in the areas of burnt tissue, blur or other tissue processing or scanning artifacts.
• Figure 3: Visualisation of applying nuclei segmentation and classification network trained on PanNuke to unseen whole-slide images. Top row: Cervix tissue, with a visible differentiation between tumor and other tissue types. Bottom row: Prostate tissue.
• Figure 4: Ground truth labels and segmentation masks verified by pathologists alongside model prediction for bladder tissue visual field. FCNN patch represents a 224${\times}$224 patch commonly used for training fully convolution segmentation models, right below it is a patch used by sirinukunwattana_locality_2016 for training a CNN to classify each individual nucleus.
• Figure 5: Left column: Pathologist verified nuclei dots; Middle: CNN generated segmentation masks; Right: NuClick generated segmentation masks and subsequently verified. Color dots are consistent with the legend in Figure \ref{['fig:bladder_images']}.