AI-driven analysis of digital pathology images may improve pediatric sarcoma subtyping

An artificial intelligence (AI)-based model accurately classified pediatric sarcomas using digital pathology images alone, according to results presented at the American Association for Cancer Research (AACR) Annual Meeting, held April 25–30.

Pediatric sarcomas are rare and diverse tumors that can form in various types of soft tissue, including muscle, tendons, fat, blood or lymphatic vessels, nerves, or the tissue surrounding joints. Sarcomas are classified into subtypes based on several factors, including the tissue of origin and various molecular features.

“Accurate classification of a patient’s sarcoma subtype is an important step that helps guide and optimize treatment,” said Adam Thiesen, an MD/Ph.D. candidate at UConn Health and The Jackson Laboratory in the lab of Jeffrey Chuang, Ph.D..

“Unfortunately, the heterogeneity of sarcomas makes them particularly difficult to classify, often requiring complex molecular and genetic testing, as well as external review by highly specialized pathologists who use pattern recognition skills honed through years of training to arrive at a diagnosis—resources that are not readily available in many health care settings.”

In this study, Thiesen and colleagues examined the potential of AI to accurately identify pediatric sarcoma subtypes. They used 691 digital images of pathology slides from collaborators at Massachusetts General Hospital, Yale New Haven Children’s Hospital, St. Jude Children’s Hospital, and the Children’s Oncology Group, representing nine sarcoma subtypes to train AI algorithms to recognize patterns associated with each subtype.

“By digitizing tissue pathology slides, we translated the visual data a pathologist normally studies into numerical data that a computer can analyze,” Thiesen explained. “Much like our cell phones can recognize a person’s face in photos and automatically generate an album of photos of that person, our AI-based models recognize certain tumor morphology patterns in the digitized slides and group them into diagnostic categories associated with specific sarcoma subtypes.”

Briefly, the researchers developed and applied open-source software to harmonize the images collected from different institutions to account for variation in format, staining, and magnification, among other factors. The harmonized images were then converted into small tiles before being fed into deep learning models that extracted numerical data for analysis by a novel statistical method. The statistical method generated summaries of each slide’s features, which were evaluated by the trained AI algorithms to categorize each slide as a specific subtype.

In validation experiments, the AI algorithms identified sarcoma subtypes with high accuracy, Thiesen reported. Specifically, the AI-driven models correctly distinguished between:

• Ewing sarcoma and other sarcoma types in 92.2% of cases;
• non-rhabdomyosarcoma soft tissue sarcomas and rhabdomyosarcoma soft tissue sarcomas in 93.8% of cases;
• alveolar rhabdomyosarcoma and embryonal rhabdomyosarcoma in 95.1% of cases; and
• alveolar rhabdomyosarcoma, embryonal rhabdomyosarcoma, and spindle cell rhabdomyosarcoma in 87.3% of cases.

“Our findings demonstrate that AI-based models can accurately diagnose various subtypes of pediatric sarcoma using only routine pathology images,” said Thiesen. “This AI-driven model could help provide more pediatric patients access to quick, streamlined, and highly accurate cancer diagnoses regardless of their geographic location or health care setting.

“Our models are built in such a way that new images can be added and trained with minimal computational equipment,” he added. “After the standard data processing, clinicians could theoretically use our models on their own laptops, which could vastly increase accessibility even in under-resourced settings.”

A limitation of the study was that the number of available pathology images was smaller than the researchers would have wanted for training AI algorithms. However, Thiesen noted that, given the rarity of pediatric sarcomas, their imaging dataset may be the largest multicenter collection of pediatric sarcomas to date, representing multiple subtypes, anatomical locations, and patient demographics.

“We hope that, over time, additional groups will work with us to further increase the size of this dataset,” said Thiesen.

The study was organized by surgical oncologist Jill Rubinstein, MD, Ph.D., senior research scientist at The Jackson Laboratory, and utilized software created by Sergii Domanskyi, Ph.D., associate computational scientist at The Jackson Laboratory.

An artificial intelligence (AI)-based model accurately classified pediatric sarcomas using digital pathology images alone, according to results presented at the American Association for Cancer Research (AACR) Annual Meeting, held April 25–30.

Pediatric sarcomas are rare and diverse tumors that can form in various types of soft tissue, including muscle, tendons, fat, blood or lymphatic vessels, nerves, or the tissue surrounding joints. Sarcomas are classified into subtypes based on several factors, including the tissue of origin and various molecular features.

“Accurate classification of a patient’s sarcoma subtype is an important step that helps guide and optimize treatment,” said Adam Thiesen, an MD/Ph.D. candidate at UConn Health and The Jackson Laboratory in the lab of Jeffrey Chuang, Ph.D..

“Unfortunately, the heterogeneity of sarcomas makes them particularly difficult to classify, often requiring complex molecular and genetic testing, as well as external review by highly specialized pathologists who use pattern recognition skills honed through years of training to arrive at a diagnosis—resources that are not readily available in many health care settings.”

In this study, Thiesen and colleagues examined the potential of AI to accurately identify pediatric sarcoma subtypes. They used 691 digital images of pathology slides from collaborators at Massachusetts General Hospital, Yale New Haven Children’s Hospital, St. Jude Children’s Hospital, and the Children’s Oncology Group, representing nine sarcoma subtypes to train AI algorithms to recognize patterns associated with each subtype.

“By digitizing tissue pathology slides, we translated the visual data a pathologist normally studies into numerical data that a computer can analyze,” Thiesen explained. “Much like our cell phones can recognize a person’s face in photos and automatically generate an album of photos of that person, our AI-based models recognize certain tumor morphology patterns in the digitized slides and group them into diagnostic categories associated with specific sarcoma subtypes.”

Briefly, the researchers developed and applied open-source software to harmonize the images collected from different institutions to account for variation in format, staining, and magnification, among other factors. The harmonized images were then converted into small tiles before being fed into deep learning models that extracted numerical data for analysis by a novel statistical method. The statistical method generated summaries of each slide’s features, which were evaluated by the trained AI algorithms to categorize each slide as a specific subtype.

In validation experiments, the AI algorithms identified sarcoma subtypes with high accuracy, Thiesen reported. Specifically, the AI-driven models correctly distinguished between:

• Ewing sarcoma and other sarcoma types in 92.2% of cases;
• non-rhabdomyosarcoma soft tissue sarcomas and rhabdomyosarcoma soft tissue sarcomas in 93.8% of cases;
• alveolar rhabdomyosarcoma and embryonal rhabdomyosarcoma in 95.1% of cases; and
• alveolar rhabdomyosarcoma, embryonal rhabdomyosarcoma, and spindle cell rhabdomyosarcoma in 87.3% of cases.

“Our findings demonstrate that AI-based models can accurately diagnose various subtypes of pediatric sarcoma using only routine pathology images,” said Thiesen. “This AI-driven model could help provide more pediatric patients access to quick, streamlined, and highly accurate cancer diagnoses regardless of their geographic location or health care setting.

“Our models are built in such a way that new images can be added and trained with minimal computational equipment,” he added. “After the standard data processing, clinicians could theoretically use our models on their own laptops, which could vastly increase accessibility even in under-resourced settings.”

A limitation of the study was that the number of available pathology images was smaller than the researchers would have wanted for training AI algorithms. However, Thiesen noted that, given the rarity of pediatric sarcomas, their imaging dataset may be the largest multicenter collection of pediatric sarcomas to date, representing multiple subtypes, anatomical locations, and patient demographics.

“We hope that, over time, additional groups will work with us to further increase the size of this dataset,” said Thiesen.

The study was organized by surgical oncologist Jill Rubinstein, MD, Ph.D., senior research scientist at The Jackson Laboratory, and utilized software created by Sergii Domanskyi, Ph.D., associate computational scientist at The Jackson Laboratory.