• Profile

AI in Cancer Diagnosis and Treatment

M3 India Newsdesk Jan 26, 2024

This article explores the transformative impact of artificial intelligence (AI) on oncology, focusing on machine learning and deep learning applications in cancer radiology, pathology, clinical oncology, and precision medicine.

Artificial intelligence (AI) is reshaping our lives in previously unimaginable ways. This is true also for oncology, where AI is now opening up new vistas for improving the management of cancer patients.

AI can be seen as a general concept indicating the ability of a machine to learn and recognise patterns and interactions from a sufficient number of representative models, and to use this information for improving the current approach towards the process of decision-making in a specific field.

Technical aspects

Two important terms, that are strictly associated with AI, should be enlightened -machine learning and deep learning.

  1. Machine learning (ML) is a general concept indicating the ability of a machine to learn and thus improve patterns and models of analysis, whereas deep learning (DL) indicates a machine-learning method that utilises complex and deep networks to finalise a highly predictive performance.
  2. DL operates by the power of multi-layered neural networks, thereby enabling self-discovery of features unknown or unanticipated by humans and eliminating manual human effort for feature extraction.
  3. Convolutional neural networks (CNNs), a type of DL, along with tremendously growing computing power have led to accelerated development of AI-based applications, particularly in medical imaging.

Current approved practice

  1. The oncology-related field that counts for the largest number of AI devices is cancer radiology, with the majority of approved devices (54.9%).
  2. It is followed by pathology (19.7%), radiation oncology (8.5%), gastroenterology (8.5%), clinical oncology (7.0%) and gynaecology 1 (1.4%).
  3. The specific tumour that counts for the largest number of AI devices is breast cancer (31.0%), followed by lung and prostate cancer (8.5% each), colorectal cancer (7.0%), brain tumours (2.8%) and others (6 types, 1.4% each).

The application of the FDA-approved devices has not been conceived as a substitute for classical analysis but is intended as an integrative tool, to be used in selected cases, potentially representing the decisive step for improving the management of cancer patients.


In the field of oncologic radiographic imaging, AI is being used for detection and diagnosis.

  1. Computer-aided detection has been used historically for breast cancer imaging, but it did not demonstrate high clinical value. Hence, breast cancer imaging has been a prime target for AI-based cancer detection.
  2. In prostate cancer, multiparametric MRI is known to increase detection of clinically relevant malignancy, but challenges such as interobserver variability remain.
  3. Detection based on AI has the potential to overcome these challenges through ML algorithms, and there are commercially available algorithms for prostate segmentation, lesion detection, and workflow integration.

Imaging models based on AI are also being used for tumour characterisation.

Characterisation may include anatomic segmentation of tumours, which allows the software to identify the borders of diseased tissue among the normal anatomy, and tumour subtype classification, which leverages clues in signal intensity, texture, shape, and other descriptors to make a diagnosis.

Segmentation—either 2D or volumetric—may be used in clinical practice for treatment decisions such as radiation planning; however, there is interobserver variability in manual tumour segmentation.

The AI-based algorithms have the potential to overcome these biases. One such product is a U.S. Food and Drug Administration–approved product that detects brain metastases and conducts segmentation for stereotactic radiosurgery.

Annotation of data from exams, such as CT imaging scans, can create volumetric units, or voxels, which are akin to 3D pixels.

These data extrapolations can lead to computer vision–based insights that are not appreciable to the naked eye. Radiomic analysis involves automated extraction of clinically relevant information from radiologic images, and it can be used to develop radiomic biomarkers through biologic validation of radiomic signatures using genetic, histologic, and other forms of correlative data.


With the accelerated growth of digital pathology, there are several applications of AI in the analysis of pathologic images for diagnosis, grading, and prognostic biomarker interpretation.

These advances have focused on :

  • Automating time-consuming tasks to increase pathologists' efficiency
  • Enabling them to increase the time they allocate to high-level
  • Decision-making tasks

This is important, particularly for those related to the complexity and confounding factors associated with disease presentation.

A DL system has been developed to assign a Gleason score, with accuracy exceeding that of general pathologists, by whole-slide image analysis of radical prostatectomy specimens.

A convolutional neural network has been used to automate the detection of tumour-infiltrating lymphocytes in images of tissue slides from The Cancer Genome Atlas, a feature prognostic of clinical outcomes for patients with 13 different cancer subtypes.

Clinical oncology

Natural language processing (NLP) is an adjacent speciality within AI that attempts to interface human language with machine interpretation; it is used to transform unstructured data—from clinical notes and diagnostic or procedural reports—into discrete data elements.

Recent advancement in the field has led to substantially increased efficacy of the technology, which can be used to automate collection and documentation of the date of diagnosis, progression-free survival, and other cancer-related tumour attributes and patient outcomes.

Such automation could support complex database- and tumour registry development, which recursively increases the power of derived models. Alone or combined with ML/DL techniques, NLP has been used for clinical trial matching and for identifying potential adverse drug reactions.

Precision oncology

The ML/DL algorithms can overcome the limitations of standard computational methods by learning patterns from the whole transcriptome. Similarly, neural networks have been applied to transcriptomic data to classify molecular subtypes of various tumours.

The application of AI to analyse large-volume multi-omics data (exome, transcriptome, and epigenome) combined with clinically annotated data sets has furthermore led to:

  • The identification of drug-susceptibility genes
  • Variant detection
  • New cancer biology insights
  • Prediction of RNA splice sites

Radiomic analysis and evolving imaging-based ML models have demonstrated the potential to predict tumour pathology and genomic alterations.

This may enable diagnosis and biomarker information without actual sampling, leading to what is called a “virtual biopsy.”

In glioblastoma, for example, non-invasive imaging-based models are being developed that can predict genetic alterations within the tumour and impact clinical management.

Limitations to the development and adoption of Artificial Intelligence models in oncology

  1. The major challenge in the development of AI models is the lack of structured, cancer-related health data, as well as the lack of standardisation in how unstructured data are collected and stored within the EHR or unified data platform of a single healthcare system.
  2. The lack of standardisation across health care systems and global communities is even more important as it limits interoperability and widespread exchange of health data and information.
  3. Acceptance of AI technologies within medicine is impeded by the ubiquitously referenced “black box” nature of the mechanism, particularly when considering DL- and neural network–based approaches, which rely on convoluted hidden layers of data interaction.
  4. Although primitive ML algorithms, like linear regression, are fully transparent in functioning, many modern approaches use strategies that involve generating many thousands of overlapping decision trees with convoluted systems of reinforcement that cannot be graphically represented to any usable degree.
  5. Interpretability is further complicated by DL, which relies on hidden layers of data interaction inspired by the interconnectedness of the neurons and synapses of the brain.
  6. Also, bias in training data sets can limit AI model transferability or result in the inability to reproduce model results in healthcare systems external to those in which they were developed and implemented.


Disclaimer- The views and opinions expressed in this article are those of the author and do not necessarily reflect the official policy or position of M3 India.

About the author of this article: Dr Bipinesh Sansar, DM Medical Oncology, Associate Professor Medical Oncology at MPMMCC and HBCH, Varanasi.

Only Doctors with an M3 India account can read this article. Sign up for free or login with your existing account.
4 reasons why Doctors love M3 India
  • Exclusive Write-ups & Webinars by KOLs

  • Nonloggedininfinity icon
    Daily Quiz by specialty
  • Nonloggedinlock icon
    Paid Market Research Surveys
  • Case discussions, News & Journals' summaries
Sign-up / Log In
M3 app logo
Choose easy access to M3 India from your mobile!

M3 instruc arrow
Add M3 India to your Home screen
Tap  Chrome menu  and select "Add to Home screen" to pin the M3 India App to your Home screen