JMIR Publications

ABSTRACT

Background:

Explainable artificial intelligence (xAI) is increasingly used in medical imaging ​to enhance transparency​, clinical ​interpretability, and trust in AI​-assisted diagnostics, particularly in oncology. Evidence on how explainability is implemented, validated, and reported in cancer imaging remains fragmented.

Objective:

This scoping review aimed ​to systematically map research​ applying xAI methods to radiologic cancer imaging, summarize methodological and clinical trends, and identify persistent gaps in validation and integration.

Methods:

We conducted a structured search of PubMed and​ Scopus, initially in December 2023 following ​the PRISMA Extension for Scoping​ ​Reviews (PRISMA-ScR​). ​The search was updated to include​ studies up to December 2024 with a focus in the xAI components of the studies that were eligible. Eligible ​peer-reviewed articles​ ​using machine learning (ML​) or deep learning (DL) were analyzed. Data from 371 studies were extracted into predefined categories covering cancer type, imaging modality, AI model, xAI method, terminology, validation, code availability, and decision support system (DSS) integration. Quantitative synthesis was complemented by descriptive analysis.

Results:

Most of the studies focused on breast (30.2%), lung (23.5%), and brain(15.1%) cancer research. Other cancer types studied included prostate, thyroid, and liver cancers. The primary imaging techniques used were computed tomography (CT) (37.5%) and magnetic resonance imaging (MRI) (28.0%), with ultrasound (US) and mammography (MMG) commonly combined for breast cancer research. DL was the methodological choice in 70.1% of studies, while classical ML accounted for 18.1%, and hybrid pipeline methods for 10%. Emerging approaches such as concept/prototype-based or causal designs accounted for 1.9% of studies. Post-hoc xAI methods were dominant (82.2%) , with visualization (53.4%) and feature relevance (36.4%) being the most common subcategories. Hybrid approaches combining post-hoc and inherent methods accounted for 12.1%, while intrinsically interpretable methods comprised 5.7% of studies. Most data used in these studies came from open/public (40.2%) or mixed (26.9%) sources. About 22.9% of studies relied on non-open institutional datasets or private datasets, and 7.8% did not report their data sources. Among validated studies, expert/user-based validation was most common (53.9%), followed by mixed methods (38.3%).. However, quantitative metrics (5.2%) and domain/clinical knowledge-based (4.1%) validation remained rare, indicating limited use of rigorous evaluation approaches. Reproducibility and translation remain challenges , with only 17.5% of studies providing code and 12.1% reporting integration into a DSS. Only a small number of studies indicated that their results were translated into actual clinical deployment.

Conclusions:

xAI is widely adopted in AI-based cancer imaging but remains methodologically inconsistent and under validated. Most research emphasizes visualization over quantitative interpretability, and few models are clinically implemented or reproducible. Future work should prioritize standardization of xAI reporting, quantitative validation of explanations, and user-centered frameworks to promote trustworthy and clinically actionable AI in oncology imaging.