New AI tool ‘TORCH’ successfully identifies cancer origins in unknown primary cases
In a recent study published in Nature Medicine, researchers developed a deep-learning approach for tumor origin differentiation using cytological histology (TORCH), recognizing malignancy and predicting tumor origin in hydrothorax and ascites using cytological pictures from 57,220 patients.
Study: Transparent medical image AI via an image–text foundation model grounded in medical literature. Image Credit: metamorworks/Shutterstock.com
Background
Cancers of unknown primary (CUP) sites are malignant illnesses diagnosed by histopathology as metastases but whose origin cannot be determined using usual diagnostic methods.
These illnesses frequently present as serous effusions and have a dismal prognosis despite combination chemotherapies. Immunohistochemistry predicts the most likely origin of CUP; however, researchers can detect a few cases using immunostaining cocktails. The accurate identification of primary sites is critical for successful and tailored therapy.
About the study
In the present study, researchers present TORCH, a deep learning algorithm, to identify cancer genesis based on cytological pictures from ascites and hydrothorax.
The researchers trained the model using four independent deep neural networks combined to produce 12 different models. Using cytological pictures, the researchers attempted to develop an artificial intelligence-based diagnostic model for predicting tumor origin among individuals with malignancy and ascites or hydrothorax metastases.
They tested and confirmed the AI system’s performance using cytological smear instances from multiple independent testing sets.
From June 2010 to October 2023, the researchers collected data from 90,572 cytological smear images from 76,183 cancer patients across four major institutions (Zhengzhou University First Hospital, Tianjin Medical University Cancer Institute and Hospital, Yantai Yuhuangding Hospital, AND Suzhou University First Hospital) as training data.
Respiratory disorders represented the highest percentage (30%, 17,058 patients) of malignant groupings.
Carcinoma accounted for 57% of ascites and hydrothorax cases, with adenocarcinoma being the most common group (47%, 27,006 patients). Only 0.6% of the squamous cell carcinomas metastasized to ascites or pleural effusion (n=346).
To test the generalizability and reliability of TORCH, the researchers included 4,520 consecutive patients from Tianjin Cancer Hospital (the Tianjin-P dataset) and 12,467 from Yantai Hospital (the Yantai dataset).
They randomly selected 496 cytology smear images from three internal testing sets to investigate whether TORCH might help junior pathologists improve their performance.
They compared the junior pathologists’ performance using TORCH to prior manual interpretation outcomes for both junior and older pathologists.
Researchers used attention heatmaps to interpret an AI model for cancer detection in 42,682 cytological smear pictures from patients at three major tertiary referral hospitals. The model was evaluated in real-world scenarios utilizing external testing datasets, which included 495 photos.
The study aims to enhance junior pathologists’ diagnostic abilities using TORCH. Ablation tests assessed the advantages of including clinical characteristics in tumor origin prediction and investigated the association between clinical factors and cytological images.
Results
The TORCH model, a novel technique for predicting tumor origins in cancer diagnosis and localization, has been evaluated on various datasets.
The findings revealed that TORCH had an overall micro-averaged one-versus-rest area under the curve (AUROC) reading of 0.97, with a top-1 accuracy of 83% and a top-3 accuracy of 99%. This enhanced TORCH’s prediction efficacy compared to pathologists, notably increasing junior pathologists’ diagnosis scores.
Patients with cancers of unknown primary whose first treatment approach was consistent with TORCH-estimated origins had a higher overall survival rate than those who received discordant therapy. The model demonstrated relatively dependable generalization and compatibility.
When coupled with five testing sets, TORCH had a top-1 accuracy of 83%, a top-2 accuracy of 96%, and a top-3 accuracy of 99%. It also produced similar micro-averaged one-versus-rest AUROC ratings in the low-certainty and high-certainty groups.
The study included 391 cancer patients, of which 276 were concordant and 115 discordant. After the follow-up period, 42% of the patients died, with 37% concordant patients and 53% discordant ones. Survival analysis revealed that concordant patients had considerably higher overall survival than discordant ones.
Poor smear preparation and image quality issues such as section folding, contaminants, or overstaining may contribute to AI overdiagnosis in pancreatic cancer. Researchers can address these flaws by meticulous manual processing throughout the data-screening step.
In the case of colonic cancer, slime took up the majority of the image’s area, which may have caused the AI model to ignore this critical aspect while reaching a diagnosis.
Conclusion
Based on the study findings, the TORCH model, an AI tool, has shown promise in clinical practice for predicting the primary system origin of malignant cells in hydrothorax and ascites.
It can distinguish between malignant tumors and benign illnesses, pinpoint cancer sources, and help in clinical decision-making in patients with cancers of unknown origin. The model performed well across five testing sets and outperformed four pathologists.
It can assist oncologists in selecting therapy for unidentified individuals with CUP, primarily adenocarcinoma, treated with empirical broad-spectrum chemotherapy regimens.
