Smarter AI For Finding Dividing Cells In Tissue Scans

Jim Crocker
14th July, 2025

Smarter AI For Finding Dividing Cells In Tissue Scans

Block diagram of the proposed mitosis detection technique.

Image adapted from: Alhassan et al. / CC BY (Source)

Key Findings

  • Researchers from Shaqra University and the University of Vermont developed a new AI model, CDL, to accurately detect cell division in cancer tissue, which is vital for diagnosis
  • The CDL model uses advanced AI techniques like transfer learning and skip connections to learn effectively from limited data and pinpoint mitotic cells with high precision
  • This new model achieved an impressive 98.8% accuracy, significantly outperforming previous methods and promising faster, more consistent cancer diagnoses
Identifying the rate at which cells divide, known as mitosis, is a critical step in diagnosing and grading many cancers, particularly breast cancer. Pathologists traditionally perform this by manually examining tissue samples under a microscope, counting mitotic cells. This manual process is not only time-consuming and tedious but can also be inconsistent, as the appearance of mitotic cells can vary, and they are often rare within a large tissue sample. This rarity creates a challenge known as "class imbalance" for automated systems, as there are far fewer examples of mitotic cells than non-mitotic cells. The need for accurate and efficient automation in this area is therefore paramount to improve cancer diagnosis. Recent research from Shaqra University and the University of Vermont introduces a new approach to this problem: a Customized Deep Learning (CDL) model[1]. This model aims to significantly improve the accuracy and reliability of automated mitosis detection. The CDL model addresses the challenges of mitosis detection by integrating several advanced deep learning techniques. Deep learning models, particularly Convolutional Neural Networks (CNNs), are a type of artificial intelligence designed to analyze visual data by learning features directly from images. They have shown great promise in medical image analysis. For instance, earlier studies have utilized deep CNNs for multi-phase mitosis detection[2] and Mask RCNN, another deep learning framework, for both detecting and outlining (segmenting) mitotic cells[3]. The CDL model builds on this foundation by refining the learning process itself. One core component of the CDL model is *transfer learning. This technique addresses the issue of class imbalance and speeds up the training process. Instead of starting from scratch, the CDL model begins with a deep learning network that has already been trained on an enormous dataset of general images (not medical). This pre-trained network has already learned to recognize fundamental visual patterns like edges, textures, and shapes. By adapting this pre-trained knowledge to the specific task of mitosis detection, the model can learn effectively even with a limited number of mitotic cell examples. This approach is similar to how other researchers have adapted state-of-the-art CNNs for cell-level discrimination through cross-domain transfer learning[2]. Another important feature of the CDL model is the use of skip connections. In complex deep learning networks, information can sometimes be lost as it passes through many layers. Skip connections provide direct pathways for information to bypass certain layers, allowing the network to retain finer details. This is particularly crucial for accurately pinpointing the exact location of mitotic cells within the complex background of tissue images. Furthermore, the CDL model incorporates an innovative selection mechanism* that combines two advanced optimization algorithms: the Jellyfish Search Optimizer (JSO) and the Walrus Optimization Algorithm (WOA). Optimization algorithms are like sophisticated guides that help the deep learning model learn more efficiently by adjusting its internal parameters during training. This hybrid approach is designed to maximize the "momentum" of the model, meaning it helps the model learn faster and converge more effectively towards the best possible solution, leading to higher accuracy. Previous research has explored various methods to improve automated mitosis detection. For example, some studies have focused on handling "weak labels," where pathologists only mark the center of a mitotic cell rather than outlining its entire shape. One such method used a "concentric loss" function to train segmentation models with these weak labels, achieving improved F-scores on challenging datasets[4]. Another study used Mask RCNN to handle both fully annotated (pixel-level) and weakly annotated (centroid-pixel) datasets, demonstrating its robustness[3]. The CDL model, while not directly focused on the weak label problem, enhances the core detection capability, achieving significantly higher performance metrics. For instance, the CDL model reports an F1 score of 0.994 and an accuracy of 98.8%. This F1 score represents a substantial improvement over previously reported figures, such as 0.75[2], 0.863[3] on the 2012 ICPR dataset, or 0.669[4] on the TUPAC16 dataset. This indicates that the CDL model's integrated approach to feature extraction, transfer learning, and optimized training has yielded a highly effective solution. The robust performance of the CDL model, evaluated on multiple public datasets, demonstrates its potential to greatly empower pathologists. By providing a highly accurate and automated tool for identifying mitotic figures, this methodology can lead to faster, more consistent, and ultimately more precise cancer diagnoses and prognoses. Future work will explore further refinements, including combining different methodologies and optimizing for real-time applications, as well as extending the CDL model to analyze other aspects of histopathological images.

MedicineHealthBiotech

References

Main Study

1) Mitosis detection in histopathological images using customized deep learning and hybrid optimization algorithms

Published 10th July, 2025

https://doi.org/10.1371/journal.pone.0327567

Related Studies

2) A multi-phase deep CNN based mitosis detection framework for breast cancer histopathological images.

https://doi.org/10.1038/s41598-021-85652-1

3) MaskMitosis: a deep learning framework for fully supervised, weakly supervised, and unsupervised mitosis detection in histopathology images.

https://doi.org/10.1007/s11517-020-02175-z

4) Weakly supervised mitosis detection in breast histopathology images using concentric loss.

https://doi.org/10.1016/j.media.2019.01.013


Related Articles

An unhandled error has occurred. Reload 🗙