CellOMaps: A Compact Representation for Robust Classification of Lung Adenocarcinoma Growth Patterns

TL;DR

This study introduces CellOMaps, a compact, nuclei-centric representation for LUAD growth pattern classification, enabling robust, generalizable discrimination across five histological patterns plus normal tissue. By detecting nuclei with HoVer-Net, encoding their types and centroids into a 3-channel map, and classifying using a dilated ResNet-50 on 448×448 tiles, the approach achieves state-of-the-art performance under patient-level evaluation and demonstrates strong generalization on external datasets. The work also presents preliminary evidence that pattern-level outputs can inform Tumor Mutational Burden (TMB) stratification, highlighting potential clinical utility. The authors provide extensive ablations, validate on multiple datasets, and discuss limitations such as reliance on nuclei detection accuracy, while making code and annotations publicly available for reproducibility and further research.

Abstract

Lung adenocarcinoma (LUAD) is a morphologically heterogeneous disease, characterized by five primary histological growth patterns. The classification of such patterns is crucial due to their direct relation to prognosis but the high subjectivity and observer variability pose a major challenge. Although several studies have developed machine learning methods for growth pattern classification, they either only report the predominant pattern per slide or lack proper evaluation. We propose a generalizable machine learning pipeline capable of classifying lung tissue into one of the five patterns or as non-tumor. The proposed pipeline's strength lies in a novel compact Cell Organization Maps (cellOMaps) representation that captures the cellular spatial patterns from Hematoxylin and Eosin whole slide images (WSIs). The proposed pipeline provides state-of-the-art performance on LUAD growth pattern classification when evaluated on both internal unseen slides and external datasets, significantly outperforming the current approaches. In addition, our preliminary results show that the model's outputs can be used to predict patients Tumor Mutational Burden (TMB) levels.

Figures (8)

• Figure 1: Sample images of lung adenocarcinoma (LUAD) growth patterns: (a) the H&E image, (b) the corresponding CellOMaps, where green, red, and blue dots denote neoplastic cells, connective cells, and non-neoplastic cells, respectively.
• Figure 2: An overview of the proposed model for growth pattern classification. (a) The formation of the CellOMaps: First, nuclei in a slide are detected and classified using HoVer-Net. Then relevant nuclei are filtered and stacked (each class in a channel) to form a 3-channel image; (b) The CellOMaps is input to a convolution neural network (ResNet-50) with an extra dilation layer, for growth pattern classification; (c) Projection of the predicted patterns on sample WSIs, give an indication that growth patterns provide insight into TMB levels, distinguishing between high and low TMB.
• Figure 3: The difference between tile-based splitting used widely in the literature (a) and the appropriate patient level splitting employed in this study (b). In (a) tiles from a single patient could be included in both the training and test set. In (b), there is a strict boundary between training and test sets and mixing of tiles is not allowed.
• Figure 4: Average and standard deviation of AUC-ROC and macro F1-scores for growth pattern classification on TCGA-LUAD using cross validation on the patient level. (CE): cross entropy, (FL): Focal loss
• Figure 5: Pathologists assessment of the misclassified tiles. Green indicates agreement with the predicted label, red indicate agreement with the ground truth, yellow indicates the appearance of mixed patterns in the tile, and gray meaning the tile expresses something other than the predicted label or ground truth.