TL;DR
Generative multimodal large language models (MLLMs) typically provide binary "yes/no" answers using a fixed, implicit decision threshold that is often poorly suited for clinical tasks. This paper introduces a simple post-hoc method to extract continuous decision scores from a single forward pass, enabling clinicians to tune sensitivity and specificity without retraining the model.
What Problem Does This Paper Address?
In medicine, the "right" answer depends on the context. A screening tool for breast cancer should prioritize sensitivity (don't miss any cases), while a tool deciding on a risky surgery should prioritize specificity (don't operate on healthy people).
Standard MLLMs, however, are generative. When asked "Does this X-ray show pneumonia?", they sample the most likely next token—usually "yes" or "no." This effectively hard-codes a decision threshold at 0.5. If the model is slightly biased toward "no," it might miss critical findings entirely at this default setting. Furthermore, the raw probability of a "yes" token is often "diluted" by the thousands of other unrelated tokens in the model's vocabulary, making it an unreliable confidence score.
Methods
The authors propose a Single-Pass Decision Score approach. Instead of letting the model generate a full text response, they look at the probability distribution of the very first token.
- Extraction: They extract the raw probabilities for the tokens "yes" and "no" ($P_{raw}(\text{yes})$ and $P_{raw}(\text{no})$).
-
Normalization: They renormalize these two values so they sum to 1: $$P_{norm}(\text{yes}) = \frac{P_{raw}(\text{yes})}{P_{raw}(\text{yes}) + P_{raw}(\text{no})}$$ This isolates the model's binary choice from the "noise" of the rest of the vocabulary.
-
Thresholding: This normalized score is used as a continuous variable. By sweeping this score from 0 to 1, they can generate Receiver Operating Characteristic (ROC) curves and select an optimal operating point based on a validation set.
The researchers compared this "Normalized" approach against using "Raw" probabilities and "Text-elicited" probabilities (where the model is prompted to explicitly write out a numerical probability, e.g., "0.85").
Datasets
The method was evaluated across six diverse medical imaging datasets:
- Chest Radiography: CheXpert and MIMIC-CXR-JPG.
- Dermatology: HAM10000 and DERM12345.
- Ophthalmology: G1020 (glaucoma detection).
- Histopathology: NCT-CRC-HE (colorectal cancer tissue classification).
Key Insights and Ideas
- The Default is Often Broken: The most striking finding was how poorly the default 0.5 threshold performed. For identifying "normal" chest X-rays (the "No Finding" class), MedGemma-27B at default settings labeled every single image as abnormal (0% sensitivity). By simply shifting the threshold to 0.005, sensitivity jumped to 94.5%.
- Normalization as Denoising: Normalizing the yes/no probabilities consistently outperformed asking the model to "think" of a probability and write it down. This suggests that MLLMs have better internal discriminative signals than they are able to express through text generation.
- Domain-Specific Fine-Tuning Matters (Mostly): MedGemma (fine-tuned for medicine) crushed the general-purpose Gemma-3 on X-ray tasks. However, surprisingly, the general models actually performed better on some dermatology datasets, highlighting that medical fine-tuning isn't a universal "level up."
- Efficiency: Because the method only requires the first decoding step, it is incredibly fast—requiring only ~50-80ms per decision.
Key Results
The normalization approach achieved the highest Area Under the ROC Curve (AUROC) on four of the six datasets.
- CheXpert: The normalized score reached an AUROC of 0.819, significantly higher than raw probabilities (0.777) or text-elicited probabilities (0.743).
- Class-Specific Performance: As shown in Figure 3, the benefit of normalization is widespread across different medical findings, though some classes (like "Pleural Other") remain challenging for current models.
The impact of threshold shifting is best visualized through ROC curves. In Figure 4, we see that the "Optimal Threshold" (orange dot) is often far from the "Default 0.5" (blue dot), drastically changing the clinical utility of the model.
Limitations and Future Work
- Logit Access: This method requires access to token-level probabilities (logits). Many commercial MLLMs (like GPT-4o or Gemini) accessed via API do not always provide these, or provide them in a limited capacity.
- Binary Limitation: The current study focuses on binary "yes/no" questions. Extending this to multi-class problems (e.g., "Which of these 5 diseases is present?") would require more complex normalization across a larger set of verbalizer tokens.
- Prompt Sensitivity: Like all LLM tasks, the results are sensitive to how the prompt is phrased. Future work could investigate how to make these decision scores more robust to prompt variations.
- Reasoning vs. Intuition: The authors note that while this "System One" (fast, intuitive) approach works well, it remains to be seen how "Chain-of-Thought" reasoning might bias or improve these underlying probabilities.
Summary updated