Deep Learning Improves CT Assessment of Temporal Bone Lesions

Combining subtraction iodine maps (SIM) with deep learning reconstruction (DLR) improved computed tomography (CT) image quality and visualisation of bone involvement in patients with temporal bone lesions, according to a prospective study.

The technique outperformed conventional contrast-enhanced CT using hybrid iterative reconstruction (HIR), with particularly strong improvements in lesions showing high contrast enhancement.

Evaluating Combined CT Imaging Techniques

Researchers assessed 77 patients with temporal bone lesions who underwent enhanced temporal bone CT. Four types of images were assessed: conventional enhanced CT reconstructed using HIR, enhanced CT using DLR, SIM combined with HIR, and SIM combined with DLR.

SIM images were generated by subtracting pre-contrast images from contrast-enhanced images and overlaying the resulting maps onto bone-window images.

Two radiologists independently assessed overall image quality, confidence in lesion detection, lesion contour delineation, and visualisation of temporal bone involvement. Quantitative measures, including image noise and contrast-to-noise ratio (CNR), were also compared.

Combined Technique Improves Lesion Visualisation

Compared with conventional enhanced CT using HIR, SIM-DLR significantly improved scores across all four qualitative assessment measures. The technique also reduced image noise and increased lesion CNR.

The researchers reported high agreement between the two radiologists, and the benefits of SIM were particularly evident in highly enhancing lesions. Compared with low-enhancement lesions, these showed greater improvements in lesion detection confidence, contour delineation, and visualisation of bone involvement.

Further Research Needed

The researchers concluded that combining SIM with DLR improved temporal bone CT image quality, anatomical delineation, and visualisation of bone involvement, particularly in hypervascular lesions. These improvements could help clinicians assess the extent of disease before surgery and plan treatment.

However, the study was conducted at a single centre and included a relatively small patient group. Researchers also noted that the performance of DLR algorithms may vary between CT manufacturers, limiting the generalisability of the findings.

Larger, multicentre prospective studies may validate the clinical reliability and wider applicability of SIM-DLR in the assessment of temporal bone lesions.

Reference:

Su T et al. Subtraction iodine maps combined with deep learning reconstruction for improved evaluation of bony invasion in temporal bone lesions. Eur Radiol.2026. DOI;10.1007/s00330-026-12904-x.