Systematic literature review assistant that maps research gaps

# Systematic Literature Review Plan: AI-Assisted Diagnostic Imaging in Radiology ## 1. Search Strategy (PRISMA-Compliant) ### Database Selection - **PubMed/MEDLINE** — primary biomedical literature - **IEEE Xplore** — technical/engineering approaches - **Scopus** — interdisciplinary coverage - **Cochrane Library** — existing systematic reviews - **Web of Science** — citation tracking ### Search String (PubMed syntax) ``` ("artificial intelligence" OR "machine learning" OR "deep learning" OR "convolutional neural network" OR "CNN") AND ("diagnostic imaging" OR "radiology" OR "medical imaging" OR "radiograph*" OR "CT scan" OR "MRI") AND ("accuracy" OR "sensitivity" OR "specificity" OR "AUC" OR "diagnostic performance" OR "clinical validation") ``` **Filters:** English language, peer-reviewed, 2018-2024, human subjects **Expected yield:** ~3,200 results before screening ### Adapted strings for IEEE and Scopus available on request (different field codes). --- ## 2. Inclusion & Exclusion Criteria ### Inclusion | # | Criterion | |---|-----------| | I1 | Studies evaluating AI/ML models for diagnostic imaging interpretation | | I2 | Reports quantitative performance metrics (sensitivity, specificity, AUC, accuracy) | | I3 | Uses datasets with confirmed ground truth (histopathology, expert consensus, or follow-up) | | I4 | Published in peer-reviewed journals or top-tier conference proceedings | | I5 | Sample size ≥ 100 images or ≥ 50 patients | ### Exclusion | # | Criterion | |---|-----------| | E1 | Review articles, editorials, commentaries, or letters (used for reference mining only) | | E2 | Studies focused solely on image segmentation without diagnostic output | | E3 | Animal or phantom-only studies | | E4 | Preprints without peer review | | E5 | Studies where AI assists workflow (scheduling, reporting) without diagnostic interpretation | --- ## 3. Quality Assessment Checklist Using a modified **QUADAS-2** (Quality Assessment of Diagnostic Accuracy Studies): - **Patient Selection:** Was a consecutive or random sample enrolled? Was case-control design avoided? Was the spectrum of disease representative? - **Index Test:** Were AI model parameters pre-specified? Was the threshold predetermined? Was the model blinded to reference standard? - **Reference Standard:** Is the reference standard likely to correctly classify the condition? Were all patients verified with the same reference standard? - **Flow & Timing:** Was there an appropriate interval between index test and reference? Did all patients receive the same reference standard? Were all patients included in the analysis? - **AI-Specific Items:** Was the training/test split clearly described? Was external validation performed? Was the dataset demographic diversity reported? Were failure modes analyzed? Score each domain as Low / High / Unclear risk of bias. --- ## 4. Data Extraction Template | Field | Description | |-------|-------------| | Study ID | First author, year, journal | | Country & Setting | Geographic location, clinical setting (academic/community) | | Imaging Modality | X-ray, CT, MRI, ultrasound, mammography, etc. | | Clinical Task | Detection, classification, staging, screening | | Body Region | Chest, brain, breast, musculoskeletal, abdomen, etc. | | AI Architecture | CNN type (ResNet, DenseNet, U-Net, Vision Transformer, etc.) | | Training Dataset Size | Number of images and patients | | Test Dataset Size | Internal vs. external validation | | Ground Truth | Histopathology, expert consensus, clinical follow-up | | Performance Metrics | Sensitivity, specificity, AUC, PPV, NPV, accuracy | | Comparator | Radiologist performance, other AI models, no comparator | | AI vs. Human Result | Superior, non-inferior, inferior, complementary | | Subgroup Analysis | By disease severity, demographics, imaging equipment | | Limitations Noted | As reported by authors | --- ## 5. Thematic Synthesis Framework **Theme 1: Technical Performance** — Pure accuracy metrics across modalities and tasks. Sub-themes: screening vs. diagnostic accuracy, single-pathology vs. multi-pathology models. **Theme 2: Clinical Integration** — Studies reporting real-world deployment, workflow impact, time savings, or reader studies with AI assistance. Sub-themes: AI as second reader, AI as triage tool, AI as standalone. **Theme 3: Generalizability** — Evidence on performance across different populations, equipment manufacturers, imaging protocols. Sub-themes: domain shift, dataset bias, demographic disparities. **Theme 4: Regulatory & Validation** — Studies addressing FDA/CE-marked tools, prospective validation, multi-site studies. Sub-themes: pre-market vs. post-market evidence, real-world performance drift. **Theme 5: Human-AI Interaction** — Evidence on how radiologists interact with AI outputs, trust calibration, automation bias, deskilling concerns. --- ## 6. Anticipated Research Gaps Based on the current landscape, expect to identify these gaps: 1. **Prospective clinical trials** — Most studies are retrospective. Fewer than 5% of published work uses prospective, real-world data. 2. **Diverse populations** — Training data is heavily skewed toward North American and European populations. Performance on underrepresented demographics is rarely reported. 3. **Long-term outcome tracking** — Almost no studies follow patients to measure whether AI-assisted diagnosis improves actual clinical outcomes (survival, treatment response). 4. **Failure mode analysis** — Most papers report aggregate metrics. Systematic analysis of when and why models fail is rare. 5. **Cost-effectiveness** — Health economic evaluations of AI diagnostic tools are nearly absent from the literature. 6. **Multi-modal integration** — Few studies combine imaging AI with clinical data, lab results, or genomics for integrated diagnostic models. These gaps should form the "future directions" section of your review and may justify follow-up primary research.