A coarse-to-fine hierarchical framework for bone marrow cell recognition: integrating morphological features with class-specific augmentation and multi-perspective explainable AI

Abstract

The classification of bone marrow (BM) cell is essential for the diagnosis of many haematological disorders. Automated cytological analysis still suffers from extreme class imbalance, very high morphological similarity between cell types, and poor interpretability of most artificial intelligence (AI) models, despite advances in medical imaging and deep learning. We present an interpretable, state-of-the-art framework for BM cell classification based on a large-scale dataset with 171,374 single-cell images annotated by experts with 21 classes. Given the high severity of the class imbalance (originally 3678:1 at times), we created a new subset of 95,865 images from the overall dataset through segmentation and feature extraction. Through this process, overrepresented classes were down-sampled, while specific types of augmentation were applied to underrepresented classes to restore balance, resulting in a ratio of 1435:1. Our recursive segmentation approach, based on CMYK (Cyan, Magenta, Yellow, and Black) and HLS (Hue, Saturation, and Lightness) colour spaces, reliably identifies nucleus, cytoplasm, and whole-cell regions. From these areas, we processed 144 biologically motivated shape, colour, texture, and fractal attributes. We build a hierarchical two-stage classification model named HierEff-S2, where an EfficientNet-B4 backbone assigns each cell to one of six morphological groups. Then, group-specific EfficientNet-B3 models perform fine-grained classification within each group. With 21 classes, this architecture obtains 86.1% accuracy and outperforms other models, including VisionMamba, Ensemble Model, and MobileNet. To promote clinical interpretability, we combine two explainable AI methods to visually highlight cell regions that lead to the model predictions: Grad-CAM and LIME. Using XAI, we report 95.2% correctness at the image level, thus providing biologically meaningful attention to the model.