Machine-assisted pathological recognition has been focused on supervised learning (SL) that suffers from a significant annotation bottleneck. We propose a semi-supervised learning (SSL) method based on the mean teacher architecture using 13,111 whole slide images of colorectal cancer from 8803 subjects from 13 independent centers. SSL (~3150 labeled, ~40,950 unlabeled; ~6300 labeled, ~37,800 unlabeled patches) performs significantly better than the SL. No significant difference is found between SSL (~6300 labeled, ~37,800 unlabeled) and SL (~44,100 labeled) at patch-level diagnoses (area under the curve (AUC): 0.980 ± 0.014 vs. 0.987 ± 0.008, P value = 0.134) and patient-level diagnoses (AUC: 0.974 ± 0.013 vs. 0.980 ± 0.010, P value = 0.117), which is close to human pathologists (average AUC: 0.969). The evaluation on 15,000 lung and 294,912 lymph node images also confirm SSL can achieve similar performance as that of SL with massive annotations. SSL dramatically reduces the annotations, which has great potential to effectively build expert-level pathological artificial intelligence platforms in practice.

The pathology is decisive for disease diagnosis but relies heavily on experienced pathologists. In recent years, there has been growing interest in the use of artificial intelligence in pathology (AIP) to enhance diagnostic accuracy and efficiency. However, the impressive performance of deep learning-based AIP in laboratory settings often proves challenging to replicate in clinical practice. As the data preparation is important for AIP, the paper has reviewed AIP-related studies in the PubMed database published from January 2017 to February 2022, and 118 studies were included. An in-depth analysis of data preparation methods is conducted, encompassing the acquisition of pathological tissue slides, data cleaning, screening, and subsequent digitization. Expert review, image annotation, dataset division for model training and validation are also discussed. Furthermore, we delve into the reasons behind the challenges in reproducing the high performance of AIP in clinical settings and present effective strategies to enhance AIP’s clinical performance. The robustness of AIP depends on a randomized collection of representative disease slides, incorporating rigorous quality control and screening, correction of digital discrepancies, reasonable annotation, and sufficient data volume. Digital pathology is fundamental in clinical-grade AIP, and the techniques of data standardization and weakly supervised learning methods based on whole slide image (WSI) are effective ways to overcome obstacles of performance reproduction. The key to performance reproducibility lies in having representative data, an adequate amount of labeling, and ensuring consistency across multiple centers. Digital pathology for clinical diagnosis, data standardization and the technique of WSI-based weakly supervised learning will hopefully build clinical-grade AIP.

Background: Current artificial intelligence (AI) in histopathology typically specializes on a single task, resulting in a heavy workload of collecting and labeling a sufficient number of images for each type of cancer. Heterogeneous transfer learning (HTL) is expected to alleviate the data bottlenecks and establish models with performance comparable to supervised learning (SL). Methods: An accurate source domain model was trained using 28,634 colorectal patches. Additionally, 1000 sentinel lymph node patches and 1008 breast patches were used to train two target domain models. The feature distribution difference between sentinel lymph node metastasis or breast cancer and CRC was reduced by heterogeneous domain adaptation, and the maximum mean difference between subdomains was used for knowledge transfer to achieve accurate classification across multiple cancers. Result: HTL on 1000 sentinel lymph node patches (L-HTL-1000) outperforms SL on 1000 sentinel lymph node patches (L-SL-1-1000) (average area under the curve (AUC) and standard deviation of L-HTL-1000 vs. L-SL-1-1000: 0.949 ± 0.004 vs. 0.931 ± 0.008, p value = 0.008). There is no significant difference between L-HTL-1000 and SL on 7104 patches (L-SL-2-7104) (0.949 ± 0.004 vs. 0.948 ± 0.008, p value = 0.742). Similar results are observed for breast cancer. B-HTL-1008 vs. B-SL-1-1008: 0.962 ± 0.017 vs. 0.943 ± 0.018, p value = 0.008; B-HTL-1008 vs. B-SL-2-5232: 0.962 ± 0.017 vs. 0.951 ± 0.023, p value = 0.148. Conclusions: HTL is capable of building accurate AI models for similar cancers using a small amount of data based on a large dataset for a certain type of cancer. HTL holds great promise for accelerating the development of AI in histopathology.

Skin diseases are numerous in types and high in incidence, posing a serious threat to human health. Accurately assessing the severity of skin diseases helps dermatologists in making personalized treatment decisions. However, focusing solely on the skin lesion itself and ignoring the true state of the surrounding skin can lead to distorted results. Assessing the severity of the condition should be a holistic process. Specifically, dermatologists need to compare the abnormal skin with surrounding skin to conduct the diagnosis. To imitate such diagnosis practice of dermatologists, we propose LSATrans, a Transformer based framework customized for severity scoring of skin diseases. Different from the Standard Self-Attention module, we propose the Lesion-aware Self-Attention (LSA) module. LSA can capture the visual features of both lesion and normal surrounding skin areas and include their relationship in modeling. In addition to LSA, the proposed LSATrans also introduces a contrastive learning strategy for further optimization. We first evaluated the performance of LSATrans in scar, atopic dermatitis, and psoriasis scoring tasks, and it achieved mean absolute errors of 0.5895, 0.5614, and 0.5416 respectively in these three tasks. Furthermore, we conducted additional validation of LSATrans’s performance in two distinct skin disease diagnosis tasks, where it demonstrated remarkable outcomes with AUCs of 0.9774 and 0.9801, respectively, in the classification of common skin diseases and subtypes of skin diseases. These results are better than existing methods, indicating that LSATrans is expected to become a universal, accurate and objective intelligent tool for scoring the severity of skin diseases.

WSI‐based classification often suffers from out‐of‐distribution (OOD) generalization due to the distribution mismatch between training on mixed patches from multiple WSIs and testing on individual WSIs with varying tissue compositions. This prior shift impairs model generalization and degrades performance. To address this issue, we propose two distribution alignment strategies: intra‐WSI rearrange and inter‐WSI rearrange, which, respectively, regulate patch distribution within individual WSIs and across different WSIs. These strategies are embedded into a transformer‐based multi‐instance learning (MIL) framework enabling more accurate and robust classification. Our method achieves excellent AUC scores of 0.959 and 0.963 on the CAMELYON16 and TCGA‐NSCLC datasets, respectively. Moreover, it reaches an average AUC of 0.974 in 5‐fold cross‐validation on a private CRC dataset, matching the performance of patch‐based approaches. Ablation studies further validate the effectiveness of our proposed strategies in mitigating the OOD challenge in WSI classification. Overall, these strategies enhance the robustness and accuracy of WSI‐based models in handling OOD challenges.