by Christos Evangelou, MSc, PhD – Medical Writer and Editor

In a recent publication, researchers presented a novel deep learning approach for accurate mitosis detection in digital pathology. Proof-of-concept studies showed that, despite being computationally small, this new model is efficient and outperforms larger, more complex systems in detecting dividing cancer cells.¹ The model also demonstrated interpretability and generalizability across different datasets, addressing a key challenge in the adoption of artificial intelligence (AI) in clinical settings.

“The proposed methodology can profoundly impact the workflow of pathologists and the overall field of digital pathology,” said Hammad Naveed, PhD, the corresponding author of the study. “The integration of reliable AI tools into clinical practice promotes the adoption of digital pathology, enhancing diagnostic accuracy and efficiency across the field, thereby advancing the overall practice of pathology.”

The report was published in the journal Digital Health.

Early and accurate cancer diagnosis is key to improving treatment outcomes. One crucial aspect of cancer diagnosis is the detection of mitosis — the division of cells — in tissue samples. The frequency of mitotic cells, known as the mitotic activity index, serves as a prognostic indicator in cancer diagnosis.¹ Traditionally, pathologists manually count mitotic cells under a microscope, which is time consuming and prone to variability between observers.

The detection of mitotic cells can be challenging, even for experienced pathologists. Mitotic nuclei are very small, often just 25×25 pixels in high-resolution digital pathology images.¹  They also vary in appearance depending on the phase of cell division, making them easily confused with other cell types or artifacts in the tissue.

Although AI and deep learning models have been developed to automate and potentially improve mitosis detection, Dr. Naveed explained that several challenges remain: “The primary challenges in developing a generalized mitosis detection method include distinguishing mitotic nuclei from other cells due to the different shapes and textures of mitosis phases, which complicates accurate detection. Additionally, the model must be specific to cancerous cells, making it challenging to ensure precision.”

Moreover, the rarity of mitotic cells compared to non-mitotic ones creates a significant imbalance in the data, posing challenges for AI models. According to Dr. Naveed, this imbalance in datasets leads to difficulties in training the model effectively. In addition, previous AI models faced challenges with real-world variability in tissue samples and needed more transparency for medical professionals to trust their decisions.¹

Recognizing these challenges, the research team designed a compact, efficient system specifically tailored to the nuances of mitosis detection. This approach contrasts most previous model development efforts focused on creating large and more complex models.¹

“The approach taken to address these challenges involves focusing on classification rather than object detection and employing a hybrid model that combines EfficientNet, ResNet, and dilated convolutions to enhance generalization capabilities,” stated Dr. Naveed. “This novel design allows the model to capture the subtle features of mitotic cells while remaining computationally efficient.”

The team used a deep learning-based stain normalization technique called StainNet to address the variability in tissue staining across different labs and equipment.¹ According to Dr. Naveed, this step helps the model generalize better to new, unseen data. Unlike many previous studies that framed mitosis detection as an object detection problem, this research treated it as a classification task.¹ This shift allowed for better handling of the small size of mitotic nuclei and the imbalanced nature of the dataset.

The compact model incorporates elements from EfficientNet for efficient feature extraction, ResNet for smooth information flow, and dilated convolutions for capturing larger spatial contexts without losing resolution.¹

Dr. Naveed explained that EfficientNet was chosen for its efficient compound scaling, which balances network depth, width, and resolution, providing an optimal trade-off between performance and computational efficiency. ResNet was utilized for its residual connections that help maintain gradient flow and preserve information across deep networks, enhancing feature extraction. Dilated convolutions were selected for their ability to increase the receptive field without losing resolution, allowing the model to capture larger spatial contexts and improve feature representation.¹

“This hybrid methodology was inspired by the need to improve both generalizability and interpretability of the model,” Dr. Naveed stated. “This combination leverages the strengths of each architecture, leading to better generalization and interpretability.”

The researchers employed a unique training approach, using samples from different species and tumor types to ensure the model’s ability to generalize. The team incorporated gradient-weighted class activation mapping (GradCam) to visualize which parts of an image the model focuses on when making decisions, enhancing trust and interpretability.¹

The model was trained and validated on the MiDoG’22 dataset, which includes whole slide images from various scanners, species, and tumor types. It was then tested on the TUPAC’16 dataset and a local whole slide image from a hospital without fine-tuning.

Commenting on the importance of the diversity of the MiDoG′22 dataset in the robustness of the model, Dr. Naveed said:

“This broad range of variability ensures that the model is exposed to various conditions during training, which helps it generalize better and perform robustly across different domains. The inclusion of diverse data reduces the likelihood of overfitting to specific data characteristics and enhances the model’s ability to handle real-world scenarios effectively.”

On the MiDoG’22 dataset, the model achieved an F1 score (a measure of accuracy) of 0.87, with high precision (0.92) and specificity (0.93).¹ Notably, without fine-tuning, the model performed well on the TUPAC’16 dataset, achieving an F1 score of 0.83. The model outperformed many larger, more complex models that were specifically trained on this dataset.¹

The authors emphasized the high efficiency of the model, which contains only 3.1 million parameters, compared to tens or even hundreds of millions in some competing models. Despite its smaller size, it outperformed these larger models.

“By automating the mitosis detection process, our model reduces the manual workload of pathologists, allowing them to concentrate on more complex diagnostic tasks. This automation also speeds up the diagnostic process, leading to quicker decision-making and potentially earlier treatment for patients,”

said Dr. Naveed.

Using GradCam visualizations, the team demonstrated that their model focuses on relevant cellular features when making decisions, aligning with how pathologists approach the task.¹ Furthermore, the model showed promising results when applied to a whole slide image from a local hospital, highlighting its potential for clinical use.

Dr. Naveed explained that interpretability is a crucial aspect of deep learning models in pathology because it ensures that the decision-making process of the model can be understood and validated by medical professionals.

“This is essential for gaining trust and ensuring clinical acceptance. GradCam plays a significant role in this regard by providing visual explanations of the model’s predictions,”

Dr. Naveed said. He further explained that GradCam visualizations highlight the regions in the input image that are most influential for the model’s decisions, helping pathologists verify that the model’s focus aligns with relevant clinical features. This not only validates the model’s reliability but also enhances trust and transparency in its predictions.

Although the model performed well across different datasets, the authors acknowledge that more extensive testing in diverse clinical settings is needed to fully validate its generalizability across cancer types and datasets.¹

Dr. Naveed revealed that the next steps in their research involve further enhancing the accuracy and usability of the mitosis detection methodology.

“This includes refining the model architecture and training it on more diverse and larger datasets to improve accuracy and robustness,”

he said.

Additionally, the team plans to develop user-friendly interfaces and tools to facilitate the integration of the model into clinical workflows.

“For instance, the usability of these solutions requires fine-tuning on hospital-specific data in limited-resource settings,”

Dr. Naveed noted, noting that these efforts aim to achieve broader clinical adoption and application of the model and improving its impact on digital pathology and patient care.

This study received financial support from the National Center in Big Data and Cloud Computing (NCBC), the National Center of Artificial Intelligence (NCAI), and the National University of Computer and Emerging Sciences (NUCES-FAST).

References

1. Farooq H, Saleem S, Aleem I, Iftikhar A, Sheikh UN, Naveed H. Toward interpretable and generalized mitosis detection in digital pathology using deep learning. Digit Health. 2024;10:20552076241255471. Published 2024 May 21. doi:10.1177/20552076241255471