The examination of tissue biopsy sections, specifically whole slide images (WSIs), yields a substantial amount of phenotypic data and serves as the fundamental basis for the field of cancer pathology (1). Recently, there has been significant advancement in the field of deep learning (DL) techniques (2). These methods have revolutionized the construction of diagnostic machines that exhibit a high level of accuracy. In fact, their performance in tasks related to cancer classification and diagnosis has been seen to be on par with, or even surpass, that of specialists who have undergone extensive training (3). However, to create effective deep neural network (DNN) models for cancer pathology, it has often been necessary to annotate every WSI on a pixel level using thorough ground-truth descriptions based on expert opinions (4). The utilization of slide-level labels in a weakly supervised scenario for training DNN classification models has exhibited remarkable accuracy in classifying test data. This achievement has significant implications for the implementation of adaptable mathematical systems for decision-making in clinical practice, as evidenced by previous studies (5–7).

In the context of cancer histology, DNN models do not process WSIs as single images at a time like regular images. Instead, WSIs are commonly broken into smaller units known as “tiles” that serve as input elements. Using tile-level DL characteristics, the entire WSI and tumor are classified. The Multiple Instance Learning (MIL) framework is used in most weakly supervised WSI classification algorithms to learn the slide-level label from each WSI as a “bag” of tiles. MIL models are permutation invariant, meaning WSI tiles have no specific ordering, which hinders their deployment and the weakly supervised learning paradigm (8).

The motivation behind this work is to address the limitations of current weakly supervised methods, which often overlook the spatial dependencies among WSI tiles. This oversight can lead to false positives and negatives, particularly given the heterogeneous nature of tumors. To overcome this challenge, we propose a novel and efficient hierarchical transformer model called Neighborhood Attention Transformer Multiple Instance Learning (NATMIL).

The novelty of our approach lies in the Neighborhood Attention mechanism, which localizes the Self-Attention operation to the nearest neighbors of each pixel, without relying on a predetermined window adjacent to the pixel. This updated definition permits all pixels to possess a uniform rate of attention, which would otherwise be diminished for edge pixels in zero-padded options. As the size of the neighborhood increases, neighborhood attention exhibits similarities to self-attention and can be considered equivalent to self-attention when the neighborhood reaches its maximum size. Moreover, the utilization of local attention offers the additional benefit of preserving translational equivariance, which sets it apart from blocked and window self-attention mechanisms.

We have devised a method called the Neighborhood Attention Transformer (NAT) that performs competitively. In conclusion, our most significant contributions are as follows:

By addressing the spatial dependencies among WSI tiles and introducing a novel attention mechanism, this work aims to significantly improve the accuracy and reliability of cancer pathology models.

In the conventional approach, a WSI is commonly partitioned into non-overlapping tiles of a predetermined size. These tiles are subsequently assigned a weak label, determined based on the diagnosis at the slide level, to be utilized as input for a Deep Neural Network (DNN) (9). The MIL formulation allows for the prediction of a WSI label (cancer yes/no, cancer type) to originate either from the tile predictions (5, 10–12) or from a higher-level bag representation arising from the aggregation of the tile features (8, 13–15). The former method is referred to as instance based. The latter method, which makes use of bag embeddings (8, 14), has been shown to perform better in experiments. Recent bag-embedding-based methods (16) use attention mechanisms, which give each tile a score reflecting its importance in the overall WSI-level representation. Most contemporary bag-embedding-based methods include attention mechanisms (16), which award a score to each tile indicating its relative contribution to the overall representation of the WSI. Attention scores facilitate the automated identification of sub-regions that possess significant diagnostic value and provide information for the label at the WSI level (15, 17, 18).

Different attention-based MIL models investigate WSI tissue structure in various ways. Many of them assume that the tiles are unrelated and randomly distributed, which is why they are permutation invariant. Based on this premise, a recent study (13) suggested an attention-based MIL pooling operator that can be taught to automatically compute the bag embedding as the weighted average of all tile features in the WSI. The adoption and modification of this operator have been extensive, with the inclusion of a clustering layer (15, 19, 20) to enhance the acquisition of semantically rich and distinct class-specific features. Nevertheless, operators that are permutation invariant lack the intrinsic ability to capture the structural dependencies that exist between various tiles in the input. For example, the DSMIL method [DSMIL (21)] employs a non-local operator to calculate an attention score for each tile. This value is determined by comparing the feature representation of the tile with that of a crucial tile. Recently, transformer-based designs have been introduced to examine the correlations among the various tiles of a whole-slide image (WSI). These architectures typically employ a learnable position-dependent signal to effectively integrate the spatial information of the picture (22, 23). To optimize for the classification challenge and generate attention scores while concurrently learning the positional embeddings, TransMIL (24) uses a transformer-like architecture. However, transformer-based methodologies might overlook the fundamental biological processes that regulate the spatial organization of the slide.

The Stand-Alone Self-Attention (SASA) (25) technique is considered one of the initial sliding window self-attention patterns. Its primary objective is to substitute convolutions in current convolutional neural networks (CNNs) (26). Striding the feature map extracts key-value pairs like a convolution with zero padding. While accuracy improved, the implementation had high latency despite lower theoretical cost. Sliding window attention, first used in Longformer (27) for language processing, was later used in Vision Longformer (ViL) (28). Although Longformer and ViL’s implementations differed from SASA, they were unable to grow to larger windows and models due to computational overhead. Liu et al. presented Window and Shifted Window (Swin) Attention (29), non-sliding window-based self-attention mechanisms (30) that split feature maps and apply self-attention to each partition individually. Swin Transformer is a pioneering hierarchical vision transformer. The feature maps are pyramid shaped, reducing spatial dimensionality and boosting depth. Swin’s structure is widely employed in CNNs, making it compatible with other networks for downstream tasks like detection and segmentation. At ImageNet-1K classification, Swin outscored DeiT, which utilizes a convolutional teacher. Swin Transformer is the leading approach for object detection on MS-COCO and semantic segmentation on ADE20K. To address the slowness of SASA, Vaswani et al. (31) introduced HaloNet, which employs a new blocked attention pattern. While this modification does violate translational equivariance, the benefits in terms of both performance and memory are acknowledged. Three phases make up HaloNet’s attention mechanism: blocking, haloing, and attention. Blocking input feature maps into non-overlapping subsets creates queries. Next, “haloed” nearby blocks are extracted as keys and values. Attention is then given to extracted queries and key-value pairs. A novel CNN architecture, ConvNeXt, was proposed by Liu et al. (32), inspired by models like Swin. The aforementioned models do not incorporate attention mechanisms; nevertheless, they demonstrate superior performance compared to Swin in several visual tasks.

Our Neighborhood Attention approach localizes the field of response to a window surrounding each query, eliminating the need for additional strategies like Swin’s cyclic shift. We present Neighborhood Attention Transformer, a hierarchy-based transformer-like model using this attention mechanism, and compare its performance to Swin on image classification, object detection, and semantic segmentation.

The NATMIL approach is founded on the premise that the surrounding neighborhood of a tile contains important information on the level of attention allocated to that specific tile by the model. By establishing a parallel between our framework and the process of analyzing a biopsy slide by a pathologist, one might conceptualize the act of zooming in and out of a particular sub-region as a means to comprehensively explore its broader surroundings, so enhancing our understanding of the adjacent micro-environment and tissue.

In NATMIL, the attention score of each tile is recalibrated by combining the attention scores of its surrounding tiles. Figure 1 provides an overview of the model. It may be broken down into four parts:

1. Each WSI undergoes a preprocessing step in which the tissue area is automatically segmented and divided into several smaller patches.

2. The patch and feature extraction module is composed of a series of convolutional, max pooling, and linear layers. Its purpose is to convert the initial tile input into low-dimensional feature representations. Let   H = { h 1 , h 2 , ⋯ , h i , ⋯ , h N } , where each h i ∈ R n × d . Here, d represents the embedding dimensions of a tile, n represents the number of tiles inside a WSI, and N represents the total number of WSIs.

3. An attention vector of dimension N × 1 is produced by a Neighborhood Attention mechanism with a contrastive learning block that incorporates the localizing self-attention to the nearest neighboring pixels.

4. A feature aggregator and classification layer that combines the slide-level prediction and tile-level attention scores produced by the one prior to it.

The outcomes of employing the NATMIL methodology for the classification of WSIs in the Camelyon16 and TCGA-NSCLC datasets are displayed in Tables 1 , 2 . All tests in this study evaluate the performance using three metrics: the area under the receiver operating characteristic curve (AUC), the slide-level accuracy (ACC) with a threshold of 0.5, and the macro-averaged F1 score. These processes facilitated an acceptable evaluation across multiple techniques and datasets of varying sizes (39).

The results presented in the tables are further elucidated in Figure 4 , which illustrates the relationship between the hyperparameter “ k ” and the corresponding area under the receiver operating characteristic curve (AUC) values for the Camelyon16 and TCGA-NSCLC histopathology datasets.

The figure demonstrates the impact of varying the neighborhood size “ k ” on the performance of the NATMIL model. For lower values of “ k ” (i.e., k ∈ 2 , 3 , 4 ), the model exhibits similar behavior across both datasets, performing consistently well under identical experimental conditions. This consistency is expected, as nearby tiles convey significant information regarding the risk of a tile being malignant. However, as the value of “k” increases, there is a progressive decline in the model’s performance, except for a notable improvement when “ k ” equals 8.

This observed phenomenon can be attributed to recurring patterns within tumors, occurring at intervals of approximately eight tiles. Thus, the significance of employing models capable of capturing both local adjacent information and overall trends in the biopsy is underscored. It is also noteworthy that selecting either “ k = 4 ” or “ k = 8 ” consistently yields satisfactory outcomes due to the spatial configuration of tiles and their neighboring elements, reminiscent of a grid-like topology.

NATMIL surpasses all previous MIL models in terms of accuracy and AUC on the Camelyon16 cancer dataset. Notably, within the Camelyon16 dataset, tumor cells might constitute a mere 5% of the WSI. The occurrence of tumor cells in tissue samples is frequently observed at a low frequency, especially in metastatic locations, where tumor cells are distributed among extensive areas of normal cells (40). Therefore, the NATMIL model, which utilizes a local neighborhood analysis to readjust attention coefficients, demonstrated superior efficacy in detecting medically significant, sparsely distributed malignant spots compared to alternative models. The performance of NATMIL on the Camelyon16 dataset exhibited substantial superiority over the other baselines. The NATMIL model demonstrates a statistically significant improvement of at least 1.5% in terms of AUC compared to other currently available models.

We present the experimental results of the proposed methods on CAMELYON-16 and TCGA lung cancer dataset in comparison to the following baselines methods: i) classic AB-MIL; ii) RNN-based RNN-MIL; iii) attention-based CLAM-SB, CLAM-MB; and iv) transformer-based MIL, Trans-MIL.

For CAMELYON-16, most slides contain only small portions of tumor over the whole tissue region. The proposed NATMIL methods with different features have outperformed other existing MIL methods except Trans-MIL, which used a transformer-based aggregator, while Trans-MIL is significantly larger in model size and computational complexity. The NATMIL achieves significant performance at AUC of 0.7% better than DTFT-MIL, as the model used different feature distillations.

For TCGA lung cancer, the proposed methods also achieve leading performances, with NATMIL obtaining the best AUC value of 94.2%. Due to the significantly larger tumor regions in positive slides, even RNN and DTFT-based MIL methods perform well on the TCGA lung cancer dataset resulting in less obvious superiority of the proposed methods over other existing methods. In comparison, for the much more challenging dataset CAMELYON-16, the proposed method present robustness to the situation of small portions of tumor regions in positive slides.

In the TCGA-NSCLC dataset, it was observed that NATMIL had superior performance compared to the other baselines that were taken into consideration. The max-pooling approach, which employs the max operator as an aggregation function, demonstrated superior performance compared to other methods. The remarkable efficacy of max pooling on this dataset can be attributed to the observation that tumor cells constitute approximately 80% of the WSI in the TCGA-NSCLC dataset. The probability of accurately labeling distinct malignant cells is significantly elevated.

In this paper, we present the first effective and scalable sliding window attention technique for vision, called Neighborhood Attention. The first aggregation method employs the independence assumption to provide an attention score for each tile in the picture, whereas the second uses vision transformers to produce an attention score that accounts for the correlation between tiles.

To re-adjust the estimated attention ratings based on the similarities they share, we have introduced NATMIL, a unique MIL vision transformer-based method that considers the interdependence of nearby tiles in a histopathological image. By leveraging the pathologists’ existing slide-level labeling, NATMIL improves performance, reduces their burden, and makes more data available.