New AI tool TORCH” successfully identifies causes of cancer in unknown primary cases

New AI tool TORCH” successfully identifies causes of cancer in unknown primary cases

In a recent study published in Nature Medicine, researchers developed a deep learning approach to tumor origin differentiation using cytological histology (TORCH) that detects malignancy and predicts tumor origin in hydrothorax and ascites using cytological images from 57,220 patients.

background

Cancers of unknown primary location (CUP) are malignant diseases that are diagnosed histopathologically as metastases, but whose origin cannot be determined using conventional diagnostic methods.

These diseases often manifest as serous effusions and have a poor prognosis despite combination chemotherapy. Immunohistochemistry predicts the most likely origin of CUP; However, researchers can detect some cases using immunostaining cocktails. The precise identification of primary locations is crucial for successful and tailored therapy.

About the study

In the present study, researchers present TORCH, a deep learning algorithm for identifying cancer development based on cytological images of ascites and hydrothorax.

The researchers trained the model using four independent deep neural networks, which were combined to create 12 different models. Using cytological images, the researchers attempted to develop an artificial intelligence-based diagnostic model to predict tumor development in people with malignancies and ascites or hydrothorax metastases.

They tested and confirmed the performance of the AI ​​system using cytology swab instances from multiple independent test sets.

From June 2010 to October 2023, researchers collected data from 90,572 cytology smear images from 76,183 cancer patients at four major institutions (Zhengzhou University First Hospital, Tianjin Medical University Cancer Institute and Hospital, Yantai Yuhuangding Hospital and Suzhou University First Hospital) as part of their training. Data.

Respiratory diseases represented the highest percentage (30%, 17,058 patients) of malignant groupings.

Carcinomas accounted for 57% of cases of ascites and hydrothorax, with adenocarcinomas being the most common group (47%, 27,006 patients). Only 0.6% of squamous cell carcinomas metastasized to ascites or pleural effusion (n=346).

To test the generalizability and reliability of TORCH, 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 in-house test sets to investigate whether TORCH could help young pathologists improve their performance.

They compared junior pathologists' performance using TORCH with previous manual interpretation results for both junior and senior pathologists.

Researchers used attention heatmaps to interpret an AI model for cancer detection in 42,682 cytology smear images from patients at three large tertiary referral hospitals. The model was evaluated in real-world scenarios using external test datasets that included 495 photos.

The aim of the study is to improve the diagnostic skills of young pathologists using TORCH. Ablation testing evaluated the benefits of incorporating clinical features in predicting tumor origin and examined the association between clinical factors and cytologic images.

Results

The TORCH model, a novel technique for predicting tumor origins in cancer diagnosis and localization, was evaluated on various datasets.

The results showed that TORCH had an overall micro-averaged one-versus-rest-area-under-the-curve (AUROC) value of 0.97, with a top-1 accuracy of 83% and a top-3 accuracy of 99%. This improved the prediction efficiency of TORCH compared to pathologists and especially increased the diagnostic results of young pathologists.

Patients with cancers of unknown origin whose initial treatment approach was consistent with the origins estimated by TORCH had a higher overall survival rate than those who received discordant therapy. The model showed relatively reliable generalization and compatibility.

Combined with five test sets, TORCH achieved a top-1 accuracy of 83%, a top-2 accuracy of 96%, and a top-3 accuracy of 99%. There were also similar micro-averaged one-versus-rest AUROC scores in the low and high certainty groups.

The study included 391 cancer patients, of which 276 were concordant and 115 were discordant. After the follow-up period, 42% of patients died, of which 37% were concordant patients and 53% were discordant patients. Survival analysis revealed that concordant patients had significantly higher overall survival than discordant ones.

Poor smear preparation and image quality issues such as cut creases, contamination, or overstaining can lead to AI overdiagnosis of pancreatic cancer. Researchers can address these deficiencies through careful manual processing throughout the data verification step.

In the case of colon cancer, mucus took up most of the image area, which may have caused the AI ​​model to ignore this critical aspect in the diagnosis.

Diploma

Based on the study results, 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 differentiate between malignant tumors and benign diseases, locate cancer sources, and aid in clinical decision-making in patients with cancers of unknown origin. The model performed well in five sets of tests and outperformed four pathologists.

It may assist oncologists in selecting therapy for unidentified individuals with CUP, primarily adenocarcinoma, treated with broad-spectrum empirical chemotherapy regimens.

Sources: