Kerala Cancer Crusade and Cancer Literacy Mission and IRIA Preventive Radiology National Program conducted a Webinar Series on “Kerala Cancer Crusade and Cancer Literacy Mission” in association with IRIA Kerala, Swasthi Foundation and Community Oncology, Regional Cancer Centre, Journo Med. The sixth-day webinar discussed “Panel Discussion among Multidisciplinary Teams in Cancer Screening”. The event was held on 20th October 2022.
Dr Ananya Panda, Associate Professor, AIIMS, Jodhpur works on Radiomics in Pancreatic as well as Prostatic Cancer spoke about the “Artificial Intelligence ( AI) Support Image Based Cancer Screening.
Artificial Intelligence (AI)- Supported Image Based Cancer Screening
- AI-supported diagnostic workflow triage
- AI Supported Image analysis
- Challenges and Barriers
Artificial Intelligence (AI): Artificial Intelligence is an umbrella-based term that refers to the epidemy of computers that mimics human intelligence.
Machine Learning (ML): Computer algorithms that improve with experience. The individuals can get trained and exposed to more and more data. Machine learning can be either supervised machine learning or unsupervised machine learning (ML). In supervised machine learning, the machine initiates a supervised allocation in raising itself. In unsupervised machine learning, the algorithm can train itself to detect the pattern that deducts the visibility of human eyes. In unsupervised learning, the model is automatically able to identify the patterns and improve more and more data to play better.
Deep Learning: It is a very subset of machine learning. The models contain layers of interconnected neurons, and it is often connected to convolutional neural networks. These have a wide range of applications in Radiology.
Radiomics and Artificial Intelligence (AI) are not synonymous. Radiomics is a particular field that refers to the extraction of a high number of quantitative features from medical images. This quantitative information is often visible to the human eye. Radiomics applies human application and multiple features are applied. The traditional radiomics pipeline the predefined features like size, shape, and first and second-order texture features. The radiomic features are used to create statistical models to support individualized diagnostics.
An example of classification is the particular region behind a malignant. It can be differentiated into high-grade and low-grade cancer. On the other hand, this radiomics is the newly emerged substance of AI. The traditional radiomics pipeline is automatically configured by the filled network. The segmentation and classification automatically represent deep learning known as deep radiomics. There is a particular region behind cancer if it’s low-grade cancer.
Applications of Artificial Intelligence (AI) in Cancer Screening
There are multiple applications of artificial intelligence (AI) in Cancer Screening.
How AI is applicable from a radiological perspective is presented here.
AI Diagnostic Workflow Triage
Artificial Intelligence (AI) can be used to filter normal/ negative examinations from diagnostic work lists. Escalate abnormal/high-risk examinations for specialist radiologist review (in tertiary cancer centres). It is widely validated for mammography for breast cancer screening. Removing negative/ normal cases improved reader efficiency and detection of true positive cases.
There are 7, 364 screening mammograms done. The commercial AI algorithm risk score is between 0 and 1. The lowest 60% of AI risk scores could be triaged to a “no radiologist” work stream without missing any cancer that otherwise has been screen detected.
The highest 5% of AI risk scores escalated to further imaging/ specialist review: Additional 27-35% of breast cancers were detected.
Diagnostic Workflow Triage for Breast Cancer Analysis
Tomosynthesis generates more images than a standard two-view mammogram. There is increased reading time. There are 13,309 Digital Breast Tomosynthesis (DBT) Images. AI could filter out 40% of mammograms from the reading work list. AI could act as a second virtual reader with diagnostic accuracy similar to radiologists.
The performance of AI is more accurate than radiomics performance.
Lesion Detection or Segmentation (Deep- Learning)
Lesion Classification (Deep Learning/ Machine Learning)
Example of end-end cancer detection pipeline
- Inputs: CT Images
- CNN-based lesion detection
- Classification CNN-based risk prediction
AI-supported image analysis
It is particularly useful where imaging plays a crucial role in screening or early diagnosis. Some of these cancers include Breast cancer, Lung cancer, Prostate cancer, Pancreatic cancer, HCC, Colon cancer and Gynecological cancer.
- Breast cancer can be detected by Mammography
- Lung cancer can be detected by a Low dose of lung CT
- Prostate cancer can be detected by Multiparametric MRI
- Pancreatic cancer can be detected by CT
- HCC can be detected by USG/ CT
- Colon cancer by CT Colonography
- Gynec cancers by USG/ MRI
Breast Cancer Screening
Stand-alone artificial intelligence for Breast cancer detection in Mammography. Breast cancer screening can be performed by professional radiologists.
Stand-alone AI: It is the potential solution for smaller centers without specialist radiologists.
Standalone Artificial Intelligence (AI) can be detected by eminent radiologists.
Stand-alone AI can study 2634 mammograms with 40% of cancers. A commercially available AI system ( Transpara) automatically evaluated mammograms and has a risk score between 1-10. The performance of the AI system was statistically non-inferior to that of the average of the 101 radiologists. The AI system has an AUC higher than 61.4% of the radiologists. In addition, the AI system has the performance of ⅔ rd of the study. It was a population-based screening program that evaluates over the 1 lakh mammogram and their costs. 88% of renal cancers and 45% of cancers develop between two intervention mammograms. It is the highest-risk route that is assigned which is more specific and more sensitive.
Lung Cancer Screening
Lung Cancer is common cancer with a high mortality rate. The early detection of lung cancer decreases the mortality rate by 20-30%. The low-dose lung CT (LD-CT) is recommended for screening of high-risk patients based on smoking history and age. These screening CT exams generate 1000’s images at < 1mm slice thickness. To view the image of high frequency, there is a limited number of available radiologists.
Prostate Cancer Screening/ Early Diagnosis
Prostate cancer is also common cancer but with relatively low cancer morbidity. The doctors typically don’t perform prostate cancer screening as there is a problem of overdiagnosis. When prostate cancer screening is performed, one ends up with low-created clinically insignificant cancers that may not lead to cancer-related morbidity. Patients diagnosed with prostate cancers don’t die due to prostate cancers, but they tend to die of other cancers. Population-based screening is really important for prostate cancer screening.
- To identify the high-risk populations.
- To detect clinically significant prostate cancer that causes survival benefits. Avoid unnecessary prostate biopsies.
- All of these can be worked with AI support.
This is a typical example of how AI can be used. It is possible to incorporate clinical information at the right age. The PSI of other densities or PSI audacity, white blood cell count, prostate volume and so on can be calculated using AI. Create order-based clinical information that comes with a probability of a particular patient having a prostate biopsy. AI can be used to automate prostate gland segmentation. The prostate virtue ligament is important while calculating the patient’s prostate density. If one can make AI, and do this automatically, it can be time-consuming.
The third algorithm of AI is to use calculate a significant prostate cancer that can be targeted for biopsy. In this, the DL-based probability map overlaid on MRI shows the highest probability of cancer. The algorithm can be reduced, and the probabilities go on.
There is a high probability of clinical security camps. The current sensitivity for AI-based localization is 87%- 92%. It reduces the radiologist’s reading time and increases confidence. It is considered a clinically significant cancer. There are commercially available prostate solutions delivered in the US and other European countries. Most of these solutions allow prostate segmentation, but the performance of commercially available AIs among the Indian population is unknown. The performance of AI software is highly dependent, and the quality of nutritionists also varies. These are vendor specific. The software of AI is scenic and image specific.
Pancreatic Cancer Screening
Pancreatic Cancer Screening is with high mortality rate. The early detection of cancer is less than 2 cm in size. It improves the outcome. The radiologists can miss 40-70% of cancers i.e., less than 2 cm. It may be a benefit in screening high-risk populations supplemented by AI-aided detection. The performance of the Indian population is unknown. The performance in turn depends on the quality of the MRI which is vendor specific. It has a disproportionately high mortality. The risk of pancreatic screening AI population can be defined. It identifies the high-risk population, i.e., risk 1 and risk 3. It is a new wish identified for patients above 50 years of age with a new onset of diabetes Mellitus (DM) and weight loss.
The Role of AI in Screening
There is a high risk of incidence eminent among pancreatic cancer screening. The newly developed diabetes and newly developed pancreatic cancer which possesses the high-risk group clinical indicators. The screening can be done via imaging and computed tomography (CT) and the AI detects cancer on computer tomography (CT). The sensitivity and specificity of pancreatic cancer patients can be detected. In non-expert periodic cases, pancreatic cancer cases are absent.
The changes of early pancreatic cancer can be observed in about 12 months before clinical diagnosis but are often missed on imaging. The dilation of the pancreatic duct is 12 months before the diagnosis is detected i.e., around 15%. The mass appears 9 months before the diagnosis. Peripancreatic cancer spreads around 6 months before diagnosis. It can be detected in only 50%. The need for artificial intelligence to pick up the changes that are visible to the naked eye takes time. The lead time is around 12-18 months.
Role of Radiomics
Radiomics-based machine learning models can detect pancreatic cancer on prediagnostic computed tomography scans at a substantial lead time before the clinical diagnosis. The CT scans were acquired 12 months before (i.e., between 3 months to 3 years) i.e., before a diagnosis of pancreatic cancer. The 4 ML models were tested. SVM Model: AUC: 0.98, Sensitivity: 96%, Specificity: 90% for prediagnostic vs. Normal CT.
Hepatocellular Carcinoma Screening
Hepatocellular Carcinoma Screening is quite common in India and Southeast Asia. The screening can confer the survival benefit. The government launches low-cost screening for liver cancer. To image and conquer the survival benefit, pancreatic cancer can uplift any model. The cirrhosis of any aetiology (Child-Pugh Class A, B), Hepatitis B viremia without cirrhosis. There are enriched biomarkers Serum AFP, blood tests etc which are semi-annual available at low cost, more accessible and available at low radiation.
There is multi-detected CT which detects hepatocellular carcinoma screening. The HCC detection rate in cirrhosis is less than in non-cirrhosis. There is a need for more robust AI models with higher detection rates and sensitivity for population screening. An opportunity to develop robust AI models is ongoing research.
Colorectal Cancer Screening: Colonography Screening
Similar to lung cancer screening, it differs from machine learning-based differentiation of benign and premalignant colorectal polyps detected with CT colonography in an asymptomatic screening population. A good proof of concept study.
Radiomics-based image analysis differentiated with benign and pre-malignant CT colonography-detected colorectal polyps with an AUC of 0.91, a sensitivity of 82% and a specificity of 85%.
The AUC for machine learning-based differentiation of benign vs premalignant polyps was 0.87 in the 6-9 mm size category and 0.90 in the 10mm or large size category.
Gynecologic Malignancies are understudied when compared to other cancers. There are no clearly defined risk cohorts that can be targeted for screening. There is deep learning-enabled pelvic ultrasound images for accurate diagnosis of ovarian cancers, this technique is widely in usage in China.
There is a road map for foundational research on AI in medical imaging.
- It improves the Data Quality
- It increases the Data Quantity
- It enables automated labelling
- It deals with clinical reality
- It understands how Artificial Intelligence (AI) works.
AI can automate routine, high-volume and time-consuming tasks. AI can serve as an excellent “virtual second reader” that helps in the early detection of cancers. AI can be used for image-based risk stratification. AI can drive increased efficiency for population screening health programs. It overcomes operational and developmental challenges currently. AI has its importance in Radiology and also in other Image-based modalities.