Deep learning-based automatic segmentation of oral squamous cell carcinoma in histopathological images: a comprehensive evaluation and performance analysis

INTRODUCTION

Cancer remains a major cause of mortality globally. According to the World Health Organization, oral cancer poses a serious public health challenge, responsible for over 300,000 deaths each year [1]. Oral squamous cell carcinoma (OSCC) is the most common type of oral cancer, which originates in the cells lining the mouth and throat, with a mortality rate exceeding 50% [1-4]. Timely detection and precise diagnosis are vital for enhancing patient survival rates, but identifying OSCC in histopathological images can be challenging, even for experienced pathologists [5-8]. Histopathological exa­mination of tissue samples is considered the definitive method for cancer diagnosis. However, manual evalua­tion conducted by pathologists can be labor-intensive, subjective, and susceptible to variations between different observers [9-12]. As a result, there is a pressing demand for automated tools to support pathologists in achieving accurate and efficient cancer diagnoses [13].
Recently, deep learning (DL) algorithms showed considerable promise in automating the detection and segmentation of cancer cells in histopathological images. These algorithms employed advanced neural networks to recognize patterns within the data and gene­rate predictions accordingly. When trained on extensive datasets of labelled images, they can accurately and rapidly identify and categorize cancer cells [14-16].
One area of focus for DL algorithms in cancer dia­gnosis is OSCC. By automatically segmenting OSCC cells in histopathological images, DL algorithms can assist pathologists in enhancing diagnostic accuracy and formulating more effective treatment strategies. How­ever, developing a DL algorithm for OSCC segmentation presents unique challenges. OSCC cells can appear in a variety of shapes and sizes, and may be surrounded by non-cancerous cells that share similar visual characteristics. Additionally, variations in image quality and staining techniques can introduce noise and other artifacts, which can interfere with the algorithm’s ability to accurately identify OSCC cells. Despite these challenges, researchers have made significant progress in developing deep learning algorithms for OSCC segmentation. By incorporating advanced techniques, such as transfer learning, data augmentation, and adversarial training, these algorithms achieved progressive results in OSCC detection and segmentation [11, 12, 17-20].
In this study, we explored the development and performance of our DL algorithm specifically designed for OSCC segmentation in histopathological images. Various techniques and approaches used by the researchers to overcome the challenges posed by OSCC segmentation were examined, and the potential implications of this technology for cancer diagnosis and treatment were discussed.

OBJECTIVES

This study aimed to develop and evaluate a DL algorithm for segmentation of OSCC in histopathological images.

MATERIAL AND METHODS

This study developed a DL algorithm for the semantic segmentation of OSCC in histopathological images. The method involves several key steps, including data preparation with specified splits for training, validation, and test sets as well as application of augmentation techniques to improve the model’s resilience, and a tailored machine-learning pipeline for precise segmentation. Performance metrics, such as Dice similarity coefficient (DSC), intersection over union (IoU), precision, and recall, were employed for the evaluation of segmentation accuracy (Figures 1 and 2).

Model pipeline

The workflow adopted in this research is described as follows: 1) pre-processing and augmenting the dataset; 2) semantically segmenting each pixel within the images; 3) precisely delineating the classified regions.
Pre-processing
During pre-processing, the dataset was randomly divided into training (85%), validation (10%), and testing (5%) sub-sets. To enhance the dataset and strengthen the model’s generalization across different data scenarios, augmentation techniques, such as horizontal and vertical flipping, were applied to each image with a 50% probability.
Semantic segmentation
The machine learning task at hand is semantic segmentation, targeting precise categorization of each pixel as either carcinoma or non-carcinoma tissue. To address this challenge, U^2-Net architecture was implemented. U^2-Net incorporates both contraction and expansion paths, capturing detailed features and context, which are instrumental in executing the pixel-wise segmentation. This model captures multi-scale features, and was pro­ven effective in dense prediction tasks.
Implementation
An advanced DL model, U^2-Net, was employed in this study. Training of the model utilized a batch size of 4, with images of dimensions 512 × 1,024 pixels and containing three color channels. The network optimization was carried out using the Adam optimizer, with an initial learning rate set at 0.0002. L2 regularization was applied with a weight decay set at 0.001. The segmentation process capitalized on a loss function, specifi­cally the sparse categorical cross-entropy. The training spanned over 250 epochs to fine-tune the network parameters for reliable predictions.

Statistical analysis

The research harnessed an array of quantitative measures to objectively assess the segmentation’s performance. Metrics, such as DSC, IoU, precision, and recall, provided a multidimensional perspective on effectiveness of the model. Specifically, these metrics determined the extent of agreement between the segmented regions by the algorithm and the expert annotations considered as the gold standard.

RESULTS

In the current comprehensive evaluation of the DL-based algorithm for OSCC segmentation in histopathological images, multiple metrics were employed to assess accuracy, robustness, and generalization capabilities. The accuracy metric revealed that the algorithm correctly classified 95.3% of OSCC pixels out of the total pixels in the images. Furthermore, DSC, measuring the overlap between the predicted OSCC regions and ground truth annotations, achieved a high value of 0.947, demonstrating the algorithm’s precise delineation of OSCC regions. IOU, indicating an overlap between the predicted and ground truth regions, reached a strong segmentation accuracy of 0.902.
The algorithm’s robust generalization to new, unseen data was evident in the maximum DSC validation of 0.862 and the maximum IoU validation of 0.77. Specifically calculated on the test dataset, the Dice coefficient was 0.865, emphasizing the algorithm’s strong performance in accurately segmenting OSCC regions in previously unseen data. The F1 score on the test dataset was 0.889, providing a balanced measure of precision and recall in OSCC segmentation, while the intersection over union on the test dataset was 0.781. Test precision and recall were notably high at 0.889 and 0.893, respectively, further confirming the algorithm’s capability to identify OSCC regions with precision. In the validation phase, the proportion of correctly identified OSCC pixels reached 85.4%, as indicated by the validation accuracy metric. The Dice coefficient validation was 0.843, underscoring the algorithm’s accurate segmentation performance, while the intersection over union on the validation dataset was 0.741, demonstrating strong segmentation accuracy.

DISCUSSION

The utilization of artificial intelligence and DL algorithms in the field of medical pathology demonstrated promising results in assisting pathologists in various organs and different ways. AI and DL algorithms were employed in tasks, such as counting mitosis and Ki67 in tumors, assessing tumor metastasis in lymph nodes, and calculating the percentage of tumor-infiltrating lymphocytes. These algorithms, including the LYNA algorithm evaluated by Yang et al. [21], showed exceptional accuracy, as indicated by area under the curve values close to 100%, signifying their potential to enhance diagnostic precision [22]. Furthermore, AI algorithms were instrumental in predicting mutation profiles from histopathological images, contributing significantly to understanding tumor sub-types and molecular characteristics. A study by Wang et al. [23] used data from the Cancer Genome Atlas and Clinical Proteomic Tumor Analysis Consortium databases to train and validate AI models, and reported promising performance in identifying genetic mutations associated with various cancers. Moreover, Elmakaty et al. [24] conducted a systematic review and meta-analysis on the precision of AI-assisted detection of OSCC, highlighting the effectiveness of AI in improving early diagnosis through various techniques. This meta-analysis revealed high accuracy in distinguishing between the healthy oral tissue and OSCC, with specific diagnostic tests, such as fluorescence spectroscopy, Raman spectroscopy, and photography, indicating distinct groups with confirmed effectiveness. Additionally, AI-assisted gene sequencing of oral tissue and machine learning assistance in conductive polymer spray mass spectrometry, showed promising results in OSCC diagnosis. Furthermore, the integration of AI into oral cancer diagnosis, particularly in conjunction with tele-cytology, was found to significantly enhance detection accuracy. Studies reported up to a 30% improvement in detection accuracy when AI was combined with tele-cytology, highlighting the synergistic potential of these technologies [24]. Oya et al. [25] conducted a study on the diagnosis of OSCC in digitized histological images using convolutional neural networks, showing the potential of AI algorithms in accurately identifying and classifying oral cancer based on histological images. Similarly, Das et al. [26] differentiated OSCC from normal tissue using DL approaches, and achieved an accuracy of 82%, surpassing other state-of-the-art models. Moreover, Yang et al. [21] demonstrated the effectiveness of a specifically designed DL model to boost both the precision and efficiency of OSCC diagnosis, underscoring the potential of AI in streamlining diagnostic processes. Additionally, Rahman et al. [27] utilized transfer learning with Alex Net in a convolutional neural network to forecast oral cancer using biopsy images of OSCC, and reported an accuracy of 90.06% based on various performance parameters. Musulin et al. [28] proposed a two-phase AI-driven system for automated grading and segmentation of OSCC from histopathological images, and reported significant performance metrics. Their integration of Xception and SWT for classification, coupled with DeepLabv3+ for segmentation, resulted in impressive outcomes, suggesting substantial potential for aiding clinicians in OSCC diagnosis by reducing variability in interpretation.
Additionally, Albalawi et al. [29] developed a deep learning model based on EfficientNetB3 architecture, designed to distinguish normal epithelium from OSCC tissues using histopathological images, and achieved an impressive accuracy of 99%. Their research underscores the promising utility of DL models in enhancing OSCC diagnosis accuracy, potentially enabling earlier detection and better patient outcomes.
Considering the increasing use of targeted therapies, AI algorithms hold promise in reducing diagnosis time and lowering the costs associated with mutation analyses. While AI has not yet completely replaced pathologists, its integration into pathology practice seems promising for enhancing both efficiency and accuracy. Collaboration between regulatory authorities and pathologists is essential to establish legal and ethical frameworks regulating the use of AI in pathology [30].
The discussed studies collectively illustrate the significant advances made in using DL algorithms to enhance diagnostic capabilities in medical pathology, particularly in the diagnosis of OSCC. These advancements not only contribute to early detection and improved patient outcomes, but also underscore the importance of continuous collaboration and regulatory oversight to guarantee the ethical and efficient incorporation of AI into pathology practices.

CONCLUSIONS

In conclusion, the application of DL algorithms in medical pathology demonstrated their potential to enhance diagnostic accuracy, improve efficiency, and reduce costs. Although AI has not fully replaced pathologists, it is believed that pathologists who utilize AI technology can perform their tasks more quickly and accurately. The incorporation of AI into diagnostic systems yielded encouraging outcomes in the early identification of OSCC as well as in predicting mutation profiles from histopathological images. However, it is crucial to consider legal and ethical regulations in the implementation of AI in pathology, and collaborative efforts between regulatory authorities and pathologists are essential in establishing appropriate frameworks.

Disclosures

1. Institutional review board statement: This study was approved by the Ethics Committee of the Saveetha Medical College and Hospital (approval number: 072/01/2024/Faculty/SRB/SMCH).
2. Assistance with the article: None.
3. Financial support and sponsorship: None.
4. Conflicts of interest: The authors declare no potential conflicts of interest concerning the research, authorship, and/or publication of this article.

References