In a recent study published in Nature Medicineresearchers developed a deep learning approach to differentiate tumor origin using cytological histology (TORCH), recognizing malignancy and predicting tumor origin in hydrothorax and ascites using cytological images from 57,220 patients.
Study: Transparent medical image AI through an image-text foundation model based on medical literature. Image credit: metamorworks/Shutterstock.com
Record
Cancers of unknown primary site (CUP) are malignant diseases diagnosed histopathologically as metastases, but their origin cannot be determined using conventional diagnostic methods.
These diseases often present as serous effusions and have a poor prognosis despite combination chemotherapy. Immunohistochemistry predicts the most likely origin of CUP. However, researchers can identify some cases using immunostaining cocktails. Accurate identification of primary sites is critical for successful and tailored treatment.
About the study
In the present study, researchers present TORCH, a deep learning algorithm, to identify cancer genesis based on cytological images of ascites and hydrothorax.
The researchers trained the model using four independent deep neural networks that were combined to produce 12 different models. Using cytological images, researchers sought to develop an AI-based diagnostic model to predict tumor origin in individuals with malignancy and ascites or hydrothorax metastases.
They tested and confirmed the performance of the AI ​​system using smear cases from multiple independent test sets.
From June 2010 to October 2023, researchers collected data from 90,572 cytological smear images from 76,183 cancer patients at four major institutions (First Hospital of Zhengzhou University, Tianjin Medical University Cancer Institute and Hospital, Yantai Yuhuangding Hospital, AND Suzhou University First Hospital). data.
Respiratory disorders accounted for the highest proportion (30%, 17,058 patients) of the malignant group.
Carcinoma accounted for 57% of ascites and hydrothorax cases, with adenocarcinoma 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 smear images from three internal test sets to investigate whether TORCH could help junior pathologists improve their performance.
They compared the performance of junior pathologists using TORCH with previous manual interpretation results for both junior and senior pathologists.
The researchers used attentional heatmaps to interpret an artificial intelligence 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, which included 495 photos.
The study aims to enhance the diagnostic skills of junior pathologists using TORCH. Catalysis trials assessed the merits of including clinical features in predicting tumor origin and investigated the correlation between clinical factors and cytological images.
Results
The TORCH model, a novel technique for tumor origin prediction in cancer diagnosis and localization, has been evaluated on various datasets.
Findings revealed that TORCH had an overall micro-mean band one versus rest under the curve (AUROC) value of 0.97, with a top-1 accuracy of 83% and a top-3 accuracy of 99%. This enhanced the predictive effectiveness of TORCH compared to pathologists, notably increasing the diagnosis scores of inferior pathologist counterparts.
Patients with cancers of unknown primary whose first treatment approach was concordant with the TORCH-assessed origin had a higher overall survival rate than those who received discordant treatment. The model showed relatively reliable generalizability and compatibility.
When combined with five test 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 mean AUROC ratings of one versus rest in the low and high certainty groups.
The study included 391 cancer patients, of whom 276 were concordant and 115 discordant. After the follow-up period, 42% of patients died, with 37% concordant patients and 53% 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 section folding, contaminants, or excessive staining may contribute to AI overdiagnosis in pancreatic cancer. Researchers can address these flaws with meticulous manual editing throughout the data review step.
In the case of colon cancer, sludge occupied the majority of the image area, which may have caused the AI ​​model to ignore this critical aspect while arriving at a diagnosis.
conclusion
Based on the study findings, the TORCH model, an artificial intelligence tool, has shown promise in clinical practice to predict the primary origin of the malignant cell system in hydrothorax and ascites.
It can distinguish between malignant tumors and benign diseases, pinpoint the sources of cancer, and help guide clinical decisions in patients with cancers of unknown origin. The model performed well on five test sets and outperformed four pathologists.
It may assist oncologists in treatment selection for unrecognized individuals with CUP, primarily adenocarcinoma, treated with empiric broad-spectrum chemotherapy regimens.
