This article introduces doctors to the potential benefits and challenges of incorporating artificial intelligence (AI) in radiology practice, including workflow optimisation, improved diagnosis accuracy, collaborative diagnostics & concerns for the responsible integration of AI in clinical practice.

Artificial intelligence (AI) is the use of computer-generated informatics which allows machines to visualise and conceptualise learning and evolve problem-solving capabilities equivalent or superior to the human brain.

The two common sub-types include:

1. Artificial narrow intelligence- where a computer performs a specific task equivalent to or better than the human mind.
2. Artificial general intelligence- where a computer goes beyond specific tasks and performs higher-order pathways to think and emulate human thought processes.

The term “Artificial Intelligence” was first used in 1956 at the summer workshop at Dartmouth College in Hanover, New Hampshire, organised by John McCarthy, an American computer scientist, pioneer, and inventor. 

Essentials

Static artificial intelligence (AI) algorithm performance will degrade over time owing to naturally occurring changes in local data and environment.

Continuous learning AI algorithms are designed to update and improve themselves as their input data, environments and targets change. These AI algorithms can “learn to learn” by discovering new features that better reflect the current data and clinical setting than when they were first developed. 

AI in Radiology

At the recent World Economic Forum, the phrase AI was replaced by IA or Intelligence Augmentation. They are merging AI with radiologists as a sort of hybrid intelligence that ensures even higher levels of analytical precision. AI has the potential to greatly improve the quality of imaging and accuracy of diagnosis and efficient treatment decisions.

1. Automated detection helps in the early diagnosis of various conditions including oncology, neurovascular and cardiac disorders. The newer protocols can assist and analyse image subsets to detect abnormalities, identify lesions and generate improved patient outcomes.
2. Deep learning modules such as the U-Net are used for improving image segmentation and data reconstruction algorithms to characterise and localise normal anatomy, anatomical variants and abnormal lesions, aiding in pre-treatment (radiotherapy/surgical) planning.
3. 3D Models are used for image segmentation with precision and these can be fused for reproduction and planning surgical protocols. There are protocols for developing detailed 3D models of anatomical structures, tumours or other abnormalities. Such immersive visualisations could enhance surgical planning, interventional procedures and radiation therapy eg orthopaedic implants in various bone tumours.
4. Quantitative analysis of human organs and extracting detailed information about a patient’s condition can be generated through various deep-learning protocols. The most common example is neuroimaging where brain volume can be assessed and alterations in brain structure can be diagnosed to analyse the connectivity pathways in monitoring various neurological disorders such as epilepsy, Alzheimer's disease motor disorders etc.

a) The workflow comprises the following steps: pre-processing of images after the acquisition, image-based clinical tasks (which usually involve the quantification of features either using engineered features with traditional machine learning or deep learning), reporting results through the generation of textual radiology reports and finally, the integration of patient information from multiple data sources.

b) AI is expected to impact image-based clinical tasks, including the detection of abnormalities; the characterisation of objects in images using segmentation, diagnosis and staging; and the monitoring of objects for diagnosis and assessment of treatment response. TNM, tumour–node–metastasis. (reproduced from Nat Rev Cancer. 2018 Aug; 18(8): 500–510.)

Advantages of incorporating AI in routine radiology

1. Superior workflow optimisation: By automated image acquisition and triaging studies based on urgency and prioritising the interpretation of critical cases. Routine reporting can reduce the workload on the reporting radiologist and improve the efficacy and accuracy of interpretation. 
2. Practice evidence-based medicine: By integrating patient data, clinical manifestations, recent medical guidelines and relevant medical literature in guiding accurate diagnosis and treatment decisions.
3. Multimodality fusion: To integrate and analyse data from multiple imaging modalities- computed tomography (CT), magnetic resonance imaging (MRI) and positron emission tomography (PET) to provide a comprehensive assessment of a patient's condition.
4. Collaborative diagnosis and consultation: Radiologists could potentially use the metaverse as a collaborative platform to discuss and diagnose complex cases with other experts worldwide. They could interact with virtual representations of patient data, including medical images, and share their expertise in real-time. This could facilitate faster and more accurate diagnoses, especially in cases where multiple perspectives are beneficial.

Few common examples used in practice

1. Classifying brain tumours in paediatric age groups with the use of MRI-based deep-learning protocols can be used to make diagnostic predictions.
2. Detecting Breast Cancer - AI can assist in the interpretation in part by identifying and characterising microcalcifications on mammograms.
3. Detecting neurological abnormalities and neurodegenerative disorders such as Alzheimer’s, Parkinson’s and amyotrophic lateral sclerosis (ALS).
4. Detection of fractures which are not easily recognised on plain X-rays.
5. Lung cancer detection- especially for solitary pulmonary nodules, where AI protocols can be used to differentiate benign from malignant nodules.

Concerns and potential pitfalls

1. Lack of standardisation: This issue makes it difficult to compare or authenticate the performance of any model. Without validation, it is difficult to decide whether a model is ready for practical implementation.
2. Lack of explainability: Deep learning algorithms work on the principles of neural network architecture and process data sets through several thousand neurons which is not comprehensible by humans to derive concrete logic behind such complex mathematics. This lack of reasoning clouds the practical use of AI models in complex matters and actual clinical integration as minute errors can have catastrophic consequences.
3. Lack of validation datasets: Extremely large patient/ imaging datasets are required to validate AI in radiology, which is a time-consuming task that hinders multiple deep learning and machine learning projects. The radiology departments should be integrated to co-develop and test various AI algorithms and provide continuous data feeds.
4. Breach of privacy: As medical researchers have access to patients' personal records for training models it is a grey area in various countries. This aspect needs a common working ground for lawmakers and policy changers and the general public to work in conjunction with people who develop and implement newer AI protocols.
5. The ethics of AI in radiology: It is a raging issue in the merging of AI with medicine. Proper data usage, data privacy and data biases are a few of the nuances of the ethics of AI.

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 Vandana Mahajan, Dept Of Radiology And PET Imaging, Apollo Cancer Hospital, Chennai Tamil Nadu.