The scan is negative. No visible tumor. The radiologist signs off. The patient's PSA is rising.
Yousefirizi et al. (arXiv:2603.19368) studied 132 prostate cancer patients with biochemical recurrence but negative PSMA PET/CT scans — the imaging equivalent of “we looked and found nothing.” They extracted radiomics features from organs prone to recurrence and found that these visually clean tissues contain measurable prognostic signal. Combined PET/CT radiomics with clinical variables achieved a C-index of 0.74, versus 0.65 for clinical variables alone.
Note (2026-11-14; read against the paper's full text): the essay says the radiomics prediction becomes more meaningful when the reader is more certain the scan is negative. The paper says the opposite: higher performance in moderate-certainty negative scans than in high-certainty scans. It also states the external validation less strongly than the essay's 'robust': it was done on 27 patients from two other centers (6 and 21) using imaging features only (clinical data were not available there), and the C-index dropped by 10.1 and 7.2 percent for PET plus CT at the two centers (the abstract says under 10 percent), which the paper describes as persistent generalization challenges. The 0.74 against 0.65 C-index is the paper's figure for 132 patients with biochemical recurrence and negative [18F]DCFPyL scans (42 progressed over a median of 25.5 months), from a post-hoc analysis using nested cross-validation; the paper says the features 'may capture' subclinical alterations. The essay's account of a shadow of disease in the texture is the essay's own.
The negative scan is not empty. It contains subclinical information below the threshold of visual detection but above the threshold of quantitative extraction. The absence of visible disease is not the absence of signal — it's the absence of signal strong enough for human pattern recognition. The machine reads what the eye dismisses.
This inverts the usual diagnostic framing. Radiomics is typically applied to visible lesions — measure the tumor, characterize its texture, predict its behavior. Here, radiomics is applied to tissue that appears normal. The features that predict progression aren't in the tumor (there is no visible tumor). They're in the microenvironment of organs where disease tends to appear — texture differences, density variations, voxel statistics that haven't yet crossed the threshold of clinical visibility.
The finding is robust on external validation but sensitive to reader diagnostic certainty. When the radiologist is more confident the scan is truly negative, the radiomics prediction becomes more meaningful — because the gap between “nothing to see” and “something to measure” widens.
The clinical implication: a negative scan is not reassurance. It's a different kind of data. The question isn't whether the scan shows disease, but whether the scan's texture contains the shadow of disease that hasn't yet declared itself.