Artificial intelligence (AI) is rapidly transforming oncology research and clinical practice, driven by three converging developments: advances in deep learning algorithms, growth in specialized computing hardware, and expanding access to large cancer datasets encompassing imaging, genomics, and clinical records. AI applications now span from fundamental biological research - such as protein folding prediction - through clinical translation, drug discovery, and healthcare delivery optimization.
This comprehensive review examines AI applications across six key cancer care domains: cancer screening and diagnosis, precision treatment planning, cancer surveillance, drug discovery, healthcare delivery, and elucidation of cancer mechanisms. The review systematically evaluates AI performance across multiple cancer types and imaging modalities, comparing AI systems to human experts and assessing clinical evidence levels.
The selection of AI models is guided by data type and clinical objective. Classical machine learning methods including logistic regression and ensemble methods analyze structured data such as genomic biomarkers and lab values for survival prediction and therapy response. Convolutional neural networks (CNNs) process imaging data including histopathology and radiology for tumor detection and grading. Transformers and recurrent neural networks process sequential genomic and clinical text data for biomarker discovery and EHR mining.
An ensemble of three deep learning models for mammographic breast cancer screening demonstrated remarkable performance across multicenter cohorts: AUC 0.889 in a UK cohort of 25,856 women and AUC 0.810 in a US cohort of 3,097 women. Compared to radiologists, the AI system achieved sensitivity improvements of +2.7% (UK) and +9.4% (US) at equivalent specificity, with strong performance maintained when the model trained on UK data was tested on US data - demonstrating cross-national generalizability.
For breast cancer detection on ultrasound, a weakly supervised deep learning model trained with only breast-level labels (no pixel-level annotations) on 288,767 exams from 143,203 patients achieved AUC 0.976 on internal testing. In a reader study comparing AI directly to radiologists, the AI achieved 94.5% sensitivity versus 90.1% for radiologists (p=0.028) and 85.6% specificity versus 80.7% (p<0.001), outperforming the average radiologist on both metrics simultaneously.
The EDL-BC system for early breast cancer detection from ultrasound demonstrated striking performance for detecting early cancers that radiologists frequently miss: sensitivity of 94.4% (internal), 100% (external cohort 1), and 80% (external cohort 2) versus only 12.1% average sensitivity for unaided radiologists alone. With AI assistance, radiologist sensitivity increased dramatically from 12.1% to 54.5%, illustrating the transformative potential of AI-assisted early detection.
The Galen Breast system (Ibex Medical Analytics), an ensemble CNN applied to H&E-stained whole slide images, demonstrated clinical-grade performance for breast pathology analysis: invasive carcinoma detection sensitivity 95.5% and specificity 93.6% (AUC 0.990), and DCIS/atypical ductal hyperplasia detection sensitivity 93.2% and specificity 93.8% (AUC 0.980). These results were validated across three cohorts including clinical deployment on 12,031 slides from 5,954 patients at Institut Curie and Maccabi Healthcare Services.
DeepGrade, an ensemble of 20 CNNs for Nottingham Histological Grade (NHG) classification on whole slide images, achieved AUC 0.919-0.937 for distinguishing NHG Grade 1 from Grade 3 tumors internally, and AUC 0.907 in external validation across 1,262 patients. Beyond standard grading, DeepGrade provides additional prognostic value by risk-stratifying Grade 2 (intermediate) tumors - the most common and prognostically ambiguous grade - into low-risk and high-risk subgroups, directly addressing a key clinical challenge.
AI analysis of 3D mammography (digital breast tomosynthesis) using a progressively trained RetinaNet demonstrated AUC values ranging from 0.927 to 0.971 across five external validation cohorts spanning the US, UK, and China, and different scanner vendors. Importantly, the model was evaluated not only on current index screening exams but also on prior negative screening exams taken 12-24 months before diagnosis, achieving +17.5% absolute sensitivity improvement in detecting early cancers missed on prior screening, suggesting AI can retrospectively identify subtle signs of developing cancer.
Colonoscopy remains the gold standard for colorectal cancer (CRC) screening, but its effectiveness varies substantially with operator expertise, contributing to variability in adenoma detection rates and interval cancers that develop from missed lesions. AI-based real-time detection systems have been developed to provide consistent, operator-independent support during colonoscopy procedures.
The CRCNet system, trained on 464,105 images from 12,179 patients, was evaluated across three independent cohorts and demonstrated sensitivity of 91.3-96.5% versus 83.8-90.3% for human endoscopists, reaching statistical significance in two of three comparisons (p<0.001 and p=0.006). The system achieved AUC values of 0.867-0.882 across external cohorts. This evidence from multicenter validation represents a high level of clinical evidence for real-time AI-assisted colonoscopy.
A real-time image recognition system for optical diagnosis - non-invasively distinguishing neoplastic from non-neoplastic colorectal polyps without biopsy - demonstrated 95.9% sensitivity for detecting neoplastic lesions and 93.3% specificity for identifying non-neoplastic lesions. A randomized controlled trial comparing autonomous AI to AI-assisted human diagnosis showed comparable performance (84.8% vs. 83.6% sensitivity), suggesting that AI may eventually support real-time decisions about whether polyps require removal.
Modern AI in oncology leverages diverse data modalities, each requiring different modeling approaches. Medical imaging data (CT, MRI, histopathology, mammography, ultrasound) is primarily analyzed using CNNs and increasingly Vision Transformers, which extract spatial features for tumor detection, segmentation, characterization, and grading. The specific architecture choice depends on task requirements, dataset size, and computational constraints.
Genomic and molecular data including DNA sequencing, RNA expression profiles, methylation patterns, and protein expression are processed using gradient boosting, random forests, and neural network-based models for tasks including molecular subtyping, biomarker discovery, and prediction of treatment response. Multi-omics integration approaches that jointly analyze data from multiple molecular layers offer improved predictive power compared to single-modality models.
Large language models (LLMs) such as GPT-based systems are increasingly applied in oncology to extract structured clinical knowledge from the vast and growing medical literature, clinical notes, and pathology reports. LLMs can accelerate hypothesis generation in cancer research by synthesizing findings across thousands of publications and identifying potentially actionable patterns in clinical text data that human researchers may overlook due to the sheer volume of information.
Despite impressive performance benchmarks, the widespread clinical integration of AI in oncology faces significant barriers. External validation is a critical but frequently missing component: many AI systems are developed and validated at single institutions using retrospective datasets, and performance often degrades when systems are deployed at new centers with different patient populations, scanner types, acquisition protocols, or clinical workflows.
Regulatory pathways for AI-based medical devices require prospective clinical evidence of safety and effectiveness, creating a higher bar than retrospective benchmarking studies. The FDA and CE marking processes for AI diagnostic tools are actively evolving to accommodate the unique characteristics of AI systems, including the potential for model performance to shift over time as patient populations and clinical practices change.
Algorithmic bias and equity represent growing concerns in clinical AI. AI systems trained predominantly on data from specific demographic groups, health systems, or geographic regions may systematically underperform for underrepresented populations, potentially exacerbating existing healthcare disparities. Addressing bias requires deliberate dataset diversification, bias auditing, and monitoring for differential performance across patient subgroups throughout the deployment lifecycle.
AI is accelerating drug discovery in oncology by enabling rapid screening of candidate compounds, predicting drug-target interactions, identifying biomarkers of treatment response, and repurposing existing drugs for new cancer indications. Deep learning models trained on molecular structure data can predict the binding affinity of candidate drugs to cancer-related proteins, dramatically reducing the experimental screening required in early drug development.
Biomarker discovery from imaging and molecular data is a particularly active application area. AI models can identify imaging features (radiomic signatures) and molecular patterns that correlate with treatment response, recurrence risk, and metastatic potential - enabling more precise patient stratification for clinical trials and treatment selection. These AI-derived biomarkers sometimes capture biological information not captured by currently established clinical biomarkers.
AI is improving precision treatment selection by integrating multi-omics data with clinical and imaging features to recommend individualized therapy regimens. For hormone receptor-positive breast cancer, AI-based multi-omics integration models can identify patients likely to benefit from chemotherapy in addition to endocrine therapy, potentially sparing patients with low-risk molecular profiles from unnecessary chemotherapy toxicity while ensuring high-risk patients receive appropriately intensive treatment.
AI has demonstrated transformative potential across the full spectrum of oncology, from early cancer detection where AI consistently outperforms average radiologists, through treatment planning, drug discovery, and patient monitoring. The convergence of improved algorithms, greater computing power, and expanding cancer datasets is enabling AI applications that address clinical challenges previously considered insurmountable.
Realizing AI's full promise requires addressing technical, regulatory, and clinical challenges in a systematic and collaborative manner. Multicenter prospective validation studies, transparent reporting of performance across demographic subgroups, and close collaboration between AI developers and clinical oncologists are essential for building the evidence base needed for regulatory approval and clinical adoption.
When applied with scientific rigor and ethical consideration, AI-driven oncology promises to accelerate research progress and improve outcomes for all cancer patients. The next generation of AI tools - combining advanced multimodal deep learning with explainability, real-world evidence monitoring, and equitable performance standards - holds the potential to fundamentally transform how cancer is detected, treated, and ultimately cured.