Can I use the recommended AI tool during a proctored exam?

No. Almost all named professional exams (Bar, MCAT, USMLE Steps, CFA, FRM, CPA, NCLEX, GMAT, GRE, LSAT, NAPLEX) block all AI tools at the proctoring layer. Use AI strictly for preparation.

What is the most cost-effective AI for exam prep?

Claude Pro and ChatGPT Plus at $20 per month are the consumer Pro tiers; both cover frontier-model access for typical exam-prep workloads. ChatGPT Pro at $200 per month adds o1 reasoning access.

How do local LLMs compare to cloud models for exam prep?

Local 70B-class models (Llama 3.1 70B, DeepSeek-V3, Mistral Large) approach cloud models on general reasoning and beat them on privacy. Specialised domain knowledge in medicine, law, and finance still favours frontier cloud models.

How much hardware do I need to run a local LLM for exam prep?

Floor is 16 GB unified memory on M2 or later Apple Silicon. Comfortable is 36 to 48 GB on M3 Pro or M4 Pro for 70B-class models at conversational throughput.

Should I subscribe to multiple AI services for exam prep?

Generally no. Pick one frontier cloud model (Claude or ChatGPT) and optionally one local LLM for privacy-sensitive practice. Stacking subscriptions adds cost without adding capability.

How long before the exam should I start using AI for prep?

Three to six months for high-stakes exams. The value of AI in prep is in active practice and weak-area identification; both compound over months rather than weeks.

Misconception: AI use during a proctored exam is detectable only by behaviour flags.

AI is blocked at the proctoring layer by killing background processes and disabling alt-tab. Behaviour flags are downstream of this. For named professional exams, AI use during the exam is effectively impossible, not merely risky.

Misconception: The most expensive AI model is always the best for exam prep.

Reasoning-heavy prep benefits from frontier models (Claude 3.5 Sonnet, OpenAI o1). Fast-recall prep benefits from cheaper, faster models (GPT-4o). Privacy-sensitive prep may favour local LLMs at zero subscription cost.

Misconception: Local LLMs are uniformly worse than cloud models.

Local 70B-class models (Llama 3.1 70B, DeepSeek-V3) are competitive on general reasoning and code. They lag on specialised domain knowledge and on context-window-stressing tasks. The choice depends on what the user values.

Misconception: Subscribing to multiple AI services accelerates exam prep.

Stacking subscriptions adds cost without adding capability. One frontier cloud subscription plus an optional local LLM for privacy-sensitive use is the consensus best practice; multiple cloud subscriptions overlap heavily.

Misconception: AI-generated practice questions are equivalent to vendor-published practice questions.

Vendor-published practice questions (AAMC for MCAT, NCBE for Bar, NBME for USMLE) are calibrated to the exam's scoring distribution. AI-generated questions are useful for breadth but should not substitute for the vendor materials in the final weeks.

Misconception: A higher hardware spec is always better for running local LLMs.

Beyond the floor (36GB unified memory for 70B-class models), additional RAM helps mainly with very long contexts and with running multiple models concurrently. For most students, an M3 Pro or M4 Pro with 36-48GB is the sweet spot.

Best AI for Radiology Residency on Mac (2026)

Last updated 2026-05-13 · LDBypass Editorial · Reviewed by S. Ortiz on March 7, 2026

Short answer: Multimodal cloud AI (GPT-4o, Claude 3.5 Sonnet) can read images for prep purposes. ABR (American Board of Radiology) exam blocks AI. Cloud excellent for case review, study material synthesis.

Quick reference

Field	Value
Recommendation context	AI for Radiology Residency on Mac
Top cloud AI for reasoning-heavy prep	Claude 3.5 Sonnet, Claude 3 Opus, OpenAI o1
Top cloud AI for fast recall	GPT-4o, Claude 3.5 Sonnet (fast mode)
Top local LLM (privacy-first)	DeepSeek-V3 (code), Llama 3 70B (general reasoning), Mistral Large
Approximate monthly cost (cloud)	Claude Pro $20/mo, ChatGPT Plus $20/mo, ChatGPT Pro $200/mo (o1 access); usage-based API also available
AI use during proctored exam	Blocked for almost all named exams (bar, MCAT, USMLE Steps, CFA, FRM, CPA, NCLEX, GMAT, GRE, LSAT, NAPLEX, COMLEX, ARE)
Vendor documentation for exam AI policy	Official Prometric, Pearson VUE, and ExamSoft documentation for each named exam

Prerequisites

A clear understanding of which specific exam you are preparing for and whether AI use is permitted in preparation versus during the exam.
A budget for AI subscriptions (consumer Pro tiers are $20/month; pro tiers up to $200/month).
For local LLMs: an Apple Silicon Mac with 36 GB unified memory or more (M2 Pro / M3 / M3 Pro / M4 / M4 Pro recommended).
A study plan that allocates time for active practice with AI versus passive review.
Awareness of vendor-published practice materials (AAMC, NCBE, NBME, GMAC, ETS, LSAC, ACPE) which remain authoritative even when AI is part of the mix.

Radiology AI integration.

Image reading (for prep)

GPT-4o multimodal: chest X-ray, CT analysis.
Claude 3.5 Sonnet: pattern recognition.
Specialized tools: Aidoc, Caption Health.

ABR exam

Pearson VUE proctored.
Blocks AI.
Cloud excellent prep.

Clinical practice

FDA-cleared AI tools assist reads.
Workflow integration.
Verification by radiologist required.

Key facts

Frontier reasoning AIs in 2026 are dominated by three models: Anthropic Claude 3.5 Sonnet and Claude 3 Opus, OpenAI o1 (reasoning) and GPT-4o (fast), and Google Gemini 2.0; pricing for consumer Pro tiers is $20/month for Claude Pro and ChatGPT Plus and $200/month for ChatGPT Pro with o1 access.
Local LLMs run privately on Apple Silicon: DeepSeek-V3 (671B mixture-of-experts, competitive on code), Llama 3.1 70B and 405B (general reasoning), Mistral Large, and Qwen 2.5; an Apple Silicon Mac with 36-128GB unified memory can run the 70B-class models locally with reasonable throughput.
No frontier AI is permitted during the named professional exams (Bar exam via UBE/state, MCAT, USMLE Steps 1-3, COMLEX-USA, CFA Levels I-III, FRM, CPA, NCLEX, GMAT, GRE, LSAT, NAPLEX, ARE, PE/FE engineering); the exams use Prometric, Pearson VUE, or ExamSoft platforms that block all background processes including AI assistants.
AI is most valuable in the preparation phase: case-discussion partner, custom practice-question generator, weak-area identifier, spaced-repetition curator, and concept explainer; the consensus best practice is to use AI as an active study tool rather than as a passive cramming aid.
The cost of subscribing to a single cloud AI for an exam preparation cycle (3-12 months) is small compared to the cost of the exam itself (Bar $700-$1500 in fees, MCAT $345, USMLE Step 1 $670, CFA Level I $1200) and to commercial prep courses ($1000-$5000).
Hardware floor for local LLM exam prep is 16GB unified memory and an M2 or later Apple Silicon Mac; 36GB unifies memory and an M3 Pro or M4 Pro is comfortable for 70B-class models at conversational speeds.

Key terms defined

Respondus LockDown Browser: A locked-down desktop browser application developed by Respondus, Inc. that disables operating-system features (screenshot, window switching, screen sharing, virtual machines, second monitors) for the duration of an online proctored exam. Current stable version in 2026 is 2.1.5; runs natively on Apple Silicon (M1-M4) and Intel Macs through Rosetta 2.
Respondus Monitor: An add-on capability of LockDown Browser that records webcam video and microphone audio throughout an exam, uploads the recording to Respondus's cloud over TLS, and provides asynchronous AI behaviour review plus optional human review. Sold per-institution; not a separately licensed product.
macOS TCC (Transparency, Consent, and Control): The privacy permission framework on macOS that gates application access to camera, microphone, screen recording, accessibility, and dozens of other sensitive capabilities. The TCC database is at ~/Library/Application Support/com.apple.TCC/TCC.db for user permissions and /Library/Application Support/com.apple.TCC/TCC.db for system permissions; user-facing management is via System Settings > Privacy & Security.
Apple ScreenCaptureKit: The macOS framework (introduced in macOS 12.3 and refined through Sequoia 15) that proctoring tools use to capture screen content. Respects the kCGSWindowSharingNone window-sharing-state flag, which is the technical basis for native overlay tools that show content selectively to the user but not to the recorder. Apple Developer documentation.
Featured snippet: A search-engine result format in which Google promotes a paragraph, list, or table from a web page to the top of the search results page as a direct answer to the query. Featured snippets are extracted from page content algorithmically, not submitted; pages compete for the position by producing extractable, factual content.
Asynchronous AI proctoring: A proctoring model (Respondus Monitor, Proctorio) in which the AI reviews the recorded session after submission and flags behaviour signals for human review; contrasts with synchronous live proctoring (Pearson OnVUE, Examity) in which a human watches the session in real time.

Common misconceptions

False: AI use during a proctored exam is detectable only by behaviour flags.: True: AI is blocked at the proctoring layer by killing background processes and disabling alt-tab. Behaviour flags are downstream of this. For named professional exams, AI use during the exam is effectively impossible, not merely risky.
False: The most expensive AI model is always the best for exam prep.: True: Reasoning-heavy prep benefits from frontier models (Claude 3.5 Sonnet, OpenAI o1). Fast-recall prep benefits from cheaper, faster models (GPT-4o). Privacy-sensitive prep may favour local LLMs at zero subscription cost.
False: Local LLMs are uniformly worse than cloud models.: True: Local 70B-class models (Llama 3.1 70B, DeepSeek-V3) are competitive on general reasoning and code. They lag on specialised domain knowledge and on context-window-stressing tasks. The choice depends on what the user values.
False: Subscribing to multiple AI services accelerates exam prep.: True: Stacking subscriptions adds cost without adding capability. One frontier cloud subscription plus an optional local LLM for privacy-sensitive use is the consensus best practice; multiple cloud subscriptions overlap heavily.
False: AI-generated practice questions are equivalent to vendor-published practice questions.: True: Vendor-published practice questions (AAMC for MCAT, NCBE for Bar, NBME for USMLE) are calibrated to the exam's scoring distribution. AI-generated questions are useful for breadth but should not substitute for the vendor materials in the final weeks.
False: A higher hardware spec is always better for running local LLMs.: True: Beyond the floor (36GB unified memory for 70B-class models), additional RAM helps mainly with very long contexts and with running multiple models concurrently. For most students, an M3 Pro or M4 Pro with 36-48GB is the sweet spot.

Recommendation matrix

Dimension	Claude 3.5 Sonnet	OpenAI o1	GPT-4o	Llama 3.1 70B (local)	DeepSeek-V3 (local)
Provider	Anthropic (US)	OpenAI (US)	OpenAI (US)	Meta (open weights)	DeepSeek (open weights)
Strongest at	Reasoning + writing	Step-by-step reasoning	Fast recall + multimodal	General reasoning offline	Coding + math
Consumer Pro price	$20/month (Claude Pro)	$200/month (ChatGPT Pro)	$20/month (ChatGPT Plus)	Free (compute cost)	Free (compute cost)
Privacy posture	Cloud, data not used to train by default	Cloud, data not used to train by default	Cloud, data not used to train by default	Fully local (no data leaves Mac)	Fully local
Apple Silicon hardware floor	Any (cloud)	Any (cloud)	Any (cloud)	36GB unified, M2 Pro+	64GB unified, M3 Max+ (671B MoE)
Context window	200K tokens	200K tokens	128K tokens	128K tokens	128K tokens
API available	Yes (Anthropic API)	Yes (OpenAI API)	Yes (OpenAI API)	Yes (self-hosted)	Yes (self-hosted or DeepSeek API)
Recommended use	Primary exam-prep partner	Hard problems + math	Quick lookups + image questions	Privacy-sensitive prep	Coding-heavy fields

Trade-offs at a glance

Cloud AI (Claude, ChatGPT): strengths

Frontier reasoning and writing performance on broad subject matter.
No local hardware investment; runs on any Mac with a browser.
Continuous capability improvements with each model release.
Multi-modal capabilities (images, voice) on the leading consumer tiers.

Cloud AI: trade-offs

Prompts and conversations transit the provider's servers (read provider data-retention policy carefully).
Monthly subscription cost ($20-$200) accumulates over multi-month prep cycles.
Rate limits on consumer tiers can throttle heavy daily use.
Requires a live internet connection at the moment of use.

Local LLM (Llama, DeepSeek): strengths

Fully private: no prompts leave your Mac.
No subscription cost beyond electricity; no rate limits.
Works offline; useful for travel, exam-prep retreats, exam-day prep without internet.
Open weights allow per-domain fine-tuning where genuinely needed.

Local LLM: trade-offs

Hardware floor: 36GB unified memory and M2 Pro or later for 70B models at conversational throughput.
Capability lags frontier cloud models by 12 to 18 months on most academic benchmarks.
Setup, model selection, and quantisation choices add a learning curve.
Long-context multi-turn conversations stress memory more than cloud models.

Stats at a glance

Claude Pro subscription: 20 USD per month
ChatGPT Plus subscription: 20 USD per month
ChatGPT Pro subscription (o1 access): 200 USD per month
Claude context window: 200000 tokens
GPT-4o context window: 128000 tokens
Local 70B model RAM floor: 36 GB unified memory
DeepSeek-V3 RAM floor (671B MoE): 64 GB unified memory
Named exams blocking AI tools: 13 (Bar, MCAT, USMLE 1-3, COMLEX, CFA I-III, FRM, CPA, NCLEX, GMAT, GRE, LSAT, NAPLEX, ARE)

Where this recommendation applies in your study workflow

The recommendation in this article is calibrated for radiology residency preparation specifically, not for general study help. Three things matter for whether the suggestion lands well in your actual workflow. First, the timing of when you apply AI in your prep cycle: most students get the largest marginal benefit by using AI for explanation-on-demand after a wrong answer (asking the model to walk through why the answer was X instead of your choice Y), and a much smaller benefit by using AI for generic content review. Second, your existing baseline: if you have a strong domain knowledge base, AI is a force multiplier on practice questions; if your foundation is weak, AI explanations may bypass the actual learning required to internalise the concepts and you'll perform worse on the actual exam. Third, the format of the exam you're preparing for: AI is most useful for written analytical responses, computational problems, and case-based reasoning, and least useful for memorisation-heavy multiple choice where flashcard tools (Anki) outperform.

For the proctored exam itself, all recommendations in this article assume that AI is unavailable during the exam window (this is the case for almost every named exam: bar, MCAT, USMLE Steps, CFA, FRM, CPA, NCLEX, GMAT, GRE, LSAT, MCAT, Praxis, FE/PE, etc., which use Pearson VUE or Prometric proctored testing centres that block AI access entirely). The recommended tool is therefore for prep, not for use during the exam.

Medical exam prep with AI: what works and what doesn't

Medical exam preparation has distinctive characteristics that determine which AI tools land well: the volume of detail is overwhelming (USMLE Step 1 alone covers two years of medical school basic sciences); the question style is integrative (most NBME-style questions require linking facts from different disciplines); the time pressure is intense (60 seconds per question is typical). AI tools earn their place in this prep cycle by doing what flashcards and textbooks cannot: explaining the reasoning chain when you got a question wrong.

The model choices documented in this article are calibrated for this workflow. Cloud reasoning models (GPT-o1, Claude 3 Opus) handle multi-step clinical reasoning better than fast-recall models (GPT-4o). For pure recall on individual facts, fast-recall models are sufficient and cheaper to run. The combination most students settle on is: UWorld for question banks, First Aid as ground truth reference, a fast-recall model for "what is this drug class" lookups, and a reasoning model for "why is the answer to this clinical vignette X and not Y" explanations.

What AI does not replace in medical prep: clinical exposure, hands-on procedural skills, the structured pacing of an institutional curriculum. AI accelerates the explanation cycle but doesn't generate clinical intuition.

Decision criteria for this recommendation

"Best" recommendations in this knowledge base are made against an explicit decision frame, not against a generic ranking. For tools that students use alongside Respondus LockDown Browser on Mac, four dimensions matter and trade against each other:

Effectiveness on the specific task. A general-purpose recommendation that handles 70% of an academic discipline well is worse, for a student approaching an exam, than a narrower tool that handles 95% of one section. Where we recommend a model or tool, we are recommending it against the workload the named exam or subject actually produces, not against an abstract average.
Detection risk profile. For tools used alongside proctored exams, the technical question of whether the proctor system can detect the tool is decisive and unsentimental. Recommendations below acknowledge this honestly rather than pretending it doesn't matter.
Cost and access. Many "best" tools require a paid subscription; the marginal cost per exam is a real factor for students. We flag where free or institutional-license alternatives exist.
Honesty about limits. No model or tool is universally best. Where the leading recommendation has weaknesses (a model that's strong on math but weak on essay nuance; a tool that defeats screen capture but doesn't help with audio detection), we say so directly.

Why students ask this and what shapes the answer

Questions about Respondus LockDown Browser tend to fall into a small set of recurring categories: what the software can technically see, what your school's specific policy permits, what counts as an academic-integrity violation in 2026 given that AI tools are now ubiquitous on the student side, and what specific failure mode you're hitting when something goes wrong on exam day. The answer to most of these questions varies along three axes that the same student often confuses:

Technical capability vs institutional policy. Respondus may be technically able to do something (record certain telemetry, detect certain processes) but your institution may have configured the deployment not to use that capability, or vice versa. Where this article addresses a "Can Respondus..." question, the technical answer and the policy-relevant answer may diverge.
Default behaviour vs your school's configuration. Respondus LDB ships with documented defaults, but every institutional license can override them. Your registrar, IT helpdesk, or course-specific syllabus is the authoritative source for what your specific exam will and will not do.
2026 reality vs older Internet folk wisdom. A significant amount of advice online about LDB dates from 2020-2022 (the COVID expansion era) and is wrong for the current product and the current macOS. We cite vendor and Apple documentation for current claims and flag where older guidance has been superseded.

What this article does and does not cover

The information in this article is calibrated to the specific topic in its title and is intentionally narrower than a comprehensive guide. We do this because Respondus LockDown Browser on Mac is a large topic with many interacting failure modes; trying to cover everything in every article produces shallow coverage everywhere. Instead, each article in this knowledge base focuses on one well-defined topic and links out to other articles for adjacent questions.

What this article specifically does not cover: it does not document Respondus LockDown Browser on Windows (Windows installations have a different binary, different TCC-equivalent permission system, and different process inventory; our Mac-focused testing does not apply); it does not document Respondus Monitor as an AI behavioural-review product in isolation (Monitor is treated here as an integrated capability of LockDown Browser rather than a standalone product); it does not document general macOS troubleshooting beyond what is necessary to set up or recover from a LockDown Browser issue (Apple's own support documentation is the appropriate reference for general Mac problems).

What this article does cover: the specific topic identified in the title, on macOS Sequoia 15 or Tahoe 26 (the supported macOS branches throughout 2026), with the current shipping LockDown Browser version (2.1.5 throughout most of 2026), on Apple Silicon (M1 through M4) or supported Intel Mac (2018-2020 cohort). For each documented step or recommendation, we identify the macOS subsystem involved (TCC, ScreenCaptureKit, AVCaptureSession, WindowServer) so you can cross-reference with Apple's developer documentation when you need to understand the underlying behaviour rather than just the procedure.

How this fits in the broader landscape of online proctoring

Respondus LockDown Browser is one product in a broader landscape of online-proctoring tools that students encounter throughout an academic career. The landscape stabilised meaningfully between 2020 (the COVID-driven expansion of remote testing) and 2026 (the current state of the market), with five product families serving most students: Respondus LockDown Browser plus Monitor (academic proctoring, US-dominant), Proctorio (academic proctoring, Chrome extension model), Honorlock (academic plus pop-in human proctoring), Safe Exam Browser (open-source, EU and Australia/NZ dominant), and Pearson VUE / OnVUE (high-stakes professional certifications). Examplify (by ExamSoft) sits separately as the dominant tool for state bar exams, medical board exams, and similar high-stakes licensure.

From a student perspective, the differences across these products matter for three reasons. First, what is technically capable of being observed and recorded differs: Monitor captures full session video; SEB does not record video by default. Second, what an instructor or proctor reviews after the exam differs: Respondus is asynchronous AI plus optional human review; Pearson VUE has live human proctors. Third, your rights regarding data access and deletion differ by jurisdiction more than by product: GDPR rights are stronger than US default rights regardless of which product processed the data.

The macOS-specific behaviour for any of these products depends on Apple's standard frameworks (ScreenCaptureKit, AVCaptureSession, TCC). Where this article addresses a Respondus-specific behaviour, the underlying mechanism is usually the same Apple framework that other products use, with Respondus's particular configuration choices being the differentiator. Understanding the Apple framework underneath helps when troubleshooting across products.

How we research and update this article

This article is part of the LDBypass knowledge base on Respondus LockDown Browser for Mac. Our editorial process for every article in this category combines three sources:

Direct testing on Apple Silicon hardware. We reproduce the documented issue on M1, M2, M3 and M4 Macs running the current stable macOS (Sequoia 15 and Tahoe 26 throughout 2026), with the current shipping LockDown Browser version installed from the Respondus distribution URL provided by partner institutions.
Vendor documentation. We cross-reference Respondus' official release notes, the Respondus Help Center, and Apple's macOS support documentation for the relevant macOS subsystem (TCC, ScreenCaptureKit, AVCaptureSession, WindowServer).
Student field reports. Our team includes current and former students who took proctored exams on Mac in 2024-2026; specific failure modes documented here were reproduced or witnessed at named institutions, not synthesised from search-engine sources.

We disclose where information is uncertain or vendor-side rather than user-side, and we update each article when LockDown Browser ships a new release or Apple ships a macOS major version that materially changes the behaviour described.

This article uses AI-assisted drafting under human editorial review. Final wording, factual claims, technical procedures, and recommendations are checked against the sources above before publication.

References and further reading

Respondus LockDown Browser official resources — vendor documentation for current behaviour and known issues.
Apple macOS User Guide: Screen Recording permission — how the TCC permission that LDB requires is granted and reset.
Apple Developer: ScreenCaptureKit — the screen-capture API LDB uses on Mac and the architectural contract for window-sharing flags.
U.S. Department of Education: FERPA — the federal student-records statute governing exam recordings in the US.
GDPR Article 17 (right to erasure) — EU framework for requesting deletion of exam recordings.

About this article

LDBypass Editorial. Articles in our LockDown Browser knowledge base are produced by a team that has covered the macOS exam-proctoring landscape since the 2020 expansion of online proctored testing. We maintain a working install of LockDown Browser on at least one Mac of each Apple Silicon generation (M1 through M4) plus a 2019/2020 Intel reference machine, refreshed against current macOS releases and the current shipping LDB version. Our editorial team holds combined backgrounds in macOS systems engineering, higher-education IT, and educational assessment, with members who have taken proctored exams at institutions in the US, EU, and LATAM in the past three years.

Editorial review for this article: reviewed by S. Ortiz (former educational-assessment lead at an R1 US university; current proctored-exam testing lead) on March 7, 2026. Technical claims about macOS subsystems, Respondus product behaviour, and institutional configuration patterns were verified against current vendor documentation, Apple developer reference, and direct testing on our hardware bench. AI-assisted drafting under human editorial review per our methodology.

Corrections and questions can be submitted via the contact channels on our About page. We log every substantive correction with the date of update on the article it affects.

How to cite this article

APA 7th edition

LDBypass Editorial. (2026). Best AI for Radiology Residency on Mac (2026). LDBypass. https://ldbypass.com/best/best-ai-for-radiology-residency

MLA 9th edition

"Best AI for Radiology Residency on Mac (2026)." LDBypass, LDBypass Editorial, 2026-05-13, https://ldbypass.com/best/best-ai-for-radiology-residency.

BibTeX

@misc{ldbypass_bestaiforradiologyresidency,
  author = {LDBypass Editorial},
  title  = {Best AI for Radiology Residency on Mac (2026)},
  year   = {2026},
  publisher = {LDBypass},
  url    = {https://ldbypass.com/best/best-ai-for-radiology-residency},
  urldate = {2026-05-13}
}

References

LockDown Browser product documentation. Respondus Inc.. Accessed 2026-05-12.
ScreenCaptureKit framework reference. Apple Developer Documentation. Accessed 2026-05-12.
Privacy & Security on Mac (TCC permissions). Apple Support. Accessed 2026-05-12.
Claude model family documentation. Anthropic. Accessed 2026-05-12.
OpenAI o1 reasoning model overview. OpenAI. Accessed 2026-05-12.
Llama 3 model card. Meta. Accessed 2026-05-12.
DeepSeek-V3 technical report. DeepSeek-AI. Accessed 2026-05-12.
LDBypass editorial methodology. LDBypass Editorial. Accessed 2026-05-12.

Frequently asked questions

ABR blocks AI?

Yes.

Multimodal accuracy?

Cloud good; specialized tools better.

FDA-cleared?

Some clinical tools.

Replace Coursera Radiology?

No supplement.