Large language models and large concept models in radiology: Present challenges, future directions, and critical perspectives

Table 1 Methodological comparison of standard large language models and large concept models

Feature	LLMs	LCMs
Level of abstraction	Token-level prediction (word/sub word)	Concept-level prediction (sentence/idea)
Input representation	Processes individual tokens, language-specific	Uses sentence embeddings, language-agnostic
Reasoning and planning	Focuses on local predictions, lacks structured reasoning	Explicitly models hierarchical reasoning and structured planning
Zero-shot generalization	Requires fine-tuning for new tasks/Languages	Strong zero-shot learning across languages and modalities
Architectural modularity	Monolithic transformer, hard to modify	Modular design, allows easy extension and updates

LLM: Large language model; LCM: Large concept model.

Table 2 Key limitations of artificial intelligence in radiology: From machine learning to large language models

Category	Description	Ref.
Data requirements	AI models (ML, DL, LLMs) require vast amounts of high-quality, annotated data, which is scarce in the medical domain. Privacy concerns and the cost of data acquisition and annotation are significant barrier	Nadkarni and Merchant[19], 2022
		Hager et al[52], 2024
Variability and bias	Differences in imaging protocols, scanner types, and patient demographics can reduce model robustness. Training on biased datasets can perpetuate and even amplify clinical disparities	Marcus et al[79], 2023
		Chen et al[80], 2021
		Guo et al[81], 2024
Incorrectness and hallucinations	LLMs, in particular, may produce outputs that are factually inaccurate or fabricated. This is a critical issue in high-stakes clinical     scenarios where accuracy is paramount	Guo et al[81], 2024
		Olabiyi et al[17], 2025
		Floridi et al[88], 2018
		Bajaj S et al[6], 2024
Limited contextual uderstanding	LLMs operate at a token level and struggle with long-range dependencies, abstract reasoning, and integrating non-linguistic data. This leads to outputs that are often superficial and lack the depth of a human physician’s diagnostic reasoning	Hendrycks et al[86], 2021
		Marcus et al[79], 2023
		Polonski et al[96], 2018
		Bender et al[87], 2021
		Najjar et al[18], 2023
		Jiang et al[57], 2023
		Nam et al[31], 2025
		Grandison[64], 2025
		Vaswani et al[30], 2017
Lack of interpretability and trust	Many AI models function as “black boxes”, providing no insight into their decision- making process. This opacity undermines trust among clinicians and poses a significant hurdle to clinical adoption and patient safety	Hendrycks et al[86], 2021
		Bender et al[87], 2021
Performance in rare diseases	AI performance often deteriorates in rare conditions, atypical presentations, and underrepresented demographics. These “edge cases” demand nuanced reasoning that goes beyond narrow pattern recognition	Busch et al[43], 2025
		Ouyang et al[84], 2025
		Atil et al[85], 2025
Obsolescence and adaptability	The rapid pace of innovation in radiology means models must continually adapt to new techniques and protocols. Without frequent retraining and validation, AI systems can become obsolete, or their performance can degrade	Najjar et al[18], 2023
Clinical integration and medicolegal	Embedding AI into clinical workflows	Nadkarni and Merchant[19], 2022
Liability	Requires extensive infrastructure and training. If an AI error leads to patient harm, the question of legal liability between developers, institutions, and clinicians remains a significant, unresolved barrier	Hager et al[52], 2024
		Soni et al[55], 2025
Computational and environmental costs	Training and running large-scale models like LLMs are resource-intensive, expensive, and have a high carbon footprint, which presents a sustainability challenge	Faiz et al[90], 2024
		The Institution of Engineering and Technology[92], 2024
		Ren et al[93], 2024
Delusions of progress	The overestimation of current AI capabilities, often due to misleading metrics, can lead to a “delusion of progress” that results in flawed decision-making and misplaced trust in high-stakes clinical scenarios	Topol[1], 2019

AI: Artificial intelligence; ML: Machine learning; DL: Deep learning; LLMs: Large language models.