Why CT slice thickness matters for lung nodules. AI-based lung nodule detection systems have shown high sensitivity (80-97%) in research datasets, but their real-world performance depends heavily on the scanning conditions and how images are reconstructed. One underexplored variable is slice thickness - how thin each CT image layer is - which directly affects how well small nodules show up.
A new generation of CT scanners. Ultra-high-resolution CT (U-HRCT) scanners can produce images with spatial resolution as fine as 0.14 mm, far exceeding conventional CT. These systems can reconstruct images as thin as 0.25 mm per slice - compared to the standard 1 mm or 5 mm used in most clinical and research settings. This opens the question of whether feeding these ultra-thin slices to AI detection systems would improve performance.
The tradeoff with thinner slices. While thinner slices theoretically capture more anatomical detail, they also increase image noise (because fewer photons hit each smaller detector element), generate vastly more images per scan, and add to radiologist reading workload. Whether these drawbacks are offset by better nodule visibility is unknown.
A practical gap in the literature. Despite the growing availability of U-HRCT and AI detection tools, no prior study had evaluated how a commercially deployed AI lung nodule detection system performs specifically on 0.25-mm U-HRCT images. This study addressed that gap directly.
Who was included. The study enrolled 63 patients (median age 74) who underwent U-HRCT for known or suspected lung cancer at a single Japanese academic hospital between July and December 2023. The majority (52 of 63) were ultimately diagnosed with primary lung cancer, while others had metastatic tumors, infections, or inflammatory conditions - a realistic mix of clinical presentations.
How the CT was performed. All patients were scanned with a Canon Aquilion Precision 160-detector U-HRCT scanner using super-high-resolution mode without contrast. Each scan was reconstructed at three different slice thicknesses from the same raw data: 5 mm (the conventional clinical standard), 1 mm (the current recommended standard for nodule detection), and 0.25 mm (the maximum resolution available).
Setting the reference standard. Two board-certified radiologists independently identified all nodules 4 mm or larger on 1-mm slice images, then reached a consensus. Importantly, any nodules that the AI detected but the radiologists had initially missed were reviewed and added to the reference list if confirmed as real - making the standard more complete than a purely human-derived list.
The AI system tested. The commercially available DL-LND system (Fujifilm SYNAPSE SAI viewer) was applied to all three slice thickness reconstructions. The system was designed to flag solid nodules larger than 3 mm and sub-solid nodules larger than 5 mm. This system is used in over 200 institutions and was originally trained on CT data acquired from 1997 onward, including the LIDC/IDRI public dataset - but not on 0.25-mm images.
Reference standard: 188 nodules. Radiologists identified 180 nodules on 1-mm slices, but the AI found 8 additional true nodules the radiologists had missed, bringing the total reference standard to 188 nodules. These included 88 solid nodules, 41 part-solid nodules, and 59 pure ground-glass nodules, with a median size of 7 mm.
1-mm slices gave the best sensitivity. The AI achieved 91.5% sensitivity on 1-mm slices - meaning it detected about 9 out of 10 real nodules. On 5-mm slices, sensitivity dropped significantly to 79.8% (p less than 0.01). On 0.25-mm slices, sensitivity was 86.7% - better than 5 mm but lower than 1 mm, with the difference between 1-mm and 0.25-mm not reaching statistical significance (p = 0.14).
More detections also means more false positives. The tradeoff for higher sensitivity on thinner slices was a lower positive predictive value (PPV) - meaning more flagged findings turned out to be false alarms. The AI's PPV was 46.4% on 5-mm slices, dropping to 34.8% on 1-mm and 36.1% on 0.25-mm slices. The mean number of false positive nodules per scan was 2.8 on 5-mm, rising to 5.2 on 1-mm and 4.7 on 0.25-mm slices.
The same pattern held for small and sub-solid nodules. For small nodules (4-10 mm), sensitivity was 72.7% on 5-mm, 89.3% on 1-mm, and 81.8% on 0.25-mm slices. For sub-solid nodules (which are particularly important to detect early), sensitivity was 69.0% on 5-mm, 87.0% on 1-mm, and 80.0% on 0.25-mm slices. In both categories, 1-mm slices outperformed 0.25-mm slices.
The AI was not trained on 0.25-mm images. The most important explanation is that the DL-LND system was trained on CT data from 1997 onward, with no 0.25-mm examples in its training library. When presented with image characteristics it has never learned to recognize, the AI system's pattern-matching breaks down. This is a fundamental limitation of how current AI systems generalize to new imaging conditions.
Some nodules became harder to see on thinner slices. Counterintuitively, ultra-thin slices can obscure certain nodule types. Ground-glass nodules - which are already faint and diffuse - were sometimes blurred or dispersed across so many ultra-thin slices that the AI could not recognize them as nodules. A specific case was shown where a 4-mm ground-glass nodule visible on 5-mm and 1-mm slices completely disappeared from AI detection on 0.25-mm slices.
Increased image noise at 0.25 mm hampers AI recognition. Thinner detector elements receive fewer X-ray photons, producing grainier images with higher noise levels. The AI system, designed to recognize nodule patterns in less noisy 1-mm images, may misinterpret noise as tissue texture or fail to recognize nodules against a noisier background at 0.25 mm.
One potential advantage of ultra-thin slices: fewer false positives. Despite lower sensitivity, 0.25-mm slices produced fewer false positive detections than 1-mm slices (4.7 vs 5.2 per scan). The higher spatial resolution helps the AI distinguish lung nodules from blood vessels, which was one common source of false alarms on thicker slices. This suggests 0.25-mm images might benefit next-generation AI systems trained on this data.
Current guideline recommendations for slice thickness. Both the Fleischner Society and the National Comprehensive Cancer Network (NCCN) recommend using slice thicknesses less than 1 mm for lung cancer screening and management of incidental pulmonary nodules, based on superior spatial resolution and reduced partial volume effects. This study's results both support and qualify those recommendations.
1 mm remains the optimal choice for current AI tools. The finding that 1-mm slices outperform both 5-mm and 0.25-mm slices for this commercially deployed system has direct practical implications. Radiology departments using U-HRCT for lung cancer screening or staging should use 1-mm reconstructions as input to AI detection tools, not push to the thinnest available setting.
Processing time is not a limiting factor. Regardless of slice thickness, the AI system processed each case in approximately 1 minute - making the computational burden of ultra-thin slices no worse than standard slices for this particular system. The bottleneck is not processing speed but detection accuracy.
AI in clinical practice still generates substantial false positive burdens. Even with the best-performing 1-mm slice setting, the average of 5.2 false positive nodule flags per scan represents a meaningful additional workload for radiologists who must review and dismiss each one. Reducing false positives without sacrificing sensitivity remains one of the central ongoing challenges for clinical AI deployment in radiology.
Single-center small sample. With 63 patients from one Japanese hospital, this study provides a first-of-its-kind evaluation but cannot be considered definitive. Results may differ in populations with different demographics, lung conditions, or cancer distributions, and with CT scanners from other manufacturers.
The reference standard was defined on 1-mm images. Because radiologists set the ground truth using 1-mm slices, there is a systematic bias: any nodules visible only at 0.25-mm resolution would not be in the reference standard. However, the researchers confirmed that no true positive nodules appeared exclusively on 5-mm or 0.25-mm slices but not on 1-mm slices, suggesting this limitation did not substantially affect the conclusions.
Matrix size limitations at 0.25 mm. Ideally, 0.25-mm slices would use 1024 or 2048 matrix reconstructions to fully exploit the spatial resolution. But the huge data volumes involved exceeded what clinical servers can handle, so the standard 512 matrix was used. This means the study tested a practical - but not theoretically maximal - representation of 0.25-mm data.
Future AI systems need training on U-HRCT data. The clearest takeaway is that next-generation AI nodule detection systems need to include ultra-high-resolution CT data in their training sets. A system trained on 0.25-mm images might leverage the superior spatial resolution to reduce false positives from vascular structures while maintaining sensitivity for true nodules - a combination that current systems cannot achieve.
The hypothesis was disproved. The researchers hypothesized that 0.25-mm slices would improve AI lung nodule detection compared to 1-mm slices. This hypothesis was not supported - 1-mm slices yielded the highest sensitivity (91.5%) and the clearest advantage over the conventional 5-mm standard (79.8%).
AI performance depends on training data compatibility. This study highlights a broader principle: AI systems perform best when the input data closely matches what they were trained on. Deploying AI in new imaging contexts - whether new CT protocols, scanner types, or slice thicknesses - requires either retraining the system or carefully validating that performance is maintained.
U-HRCT has unrealized potential for AI. The finding that 0.25-mm slices reduced certain types of false positives hints at a future where AI systems specifically trained on ultra-high-resolution data could achieve both higher sensitivity and higher specificity than currently possible. Developing and validating such systems is an important research priority.
Radiologists remain essential partners. Even with the best AI performance achieved in this study, a significant false positive rate means radiologists must carefully review all AI-flagged findings. The AI's primary value here was in flagging 8 nodules that board-certified radiologists initially missed - a clear demonstration of complementary human-AI collaboration in lung cancer detection.