The problem begins with the Response Evaluation Criteria in Solid Tumors, better known as RECIST—the internationally agreed rulebook that has governed how oncologists quantify tumor burden for decades. Under RECIST 1.1, radiologists select a small number of target lesions and track a single number for each: the longest diameter. A patient’s total tumor burden is thus compressed into a one-dimensional sum, with response categories defined by fixed thresholds, such as 30 percent shrinkage for a partial response. The scheme was celebrated for its reproducibility, but it flattens anatomy into a line. Tumors grow irregularly, cavitate, consolidate, or dissolve into scar tissue, and a diameter can remain deceptively stable while the mass behind it changes dramatically. The mismatch has grown more consequential with immunotherapy, which reshapes lesions in ways linear rules were never designed to capture. The survey highlights evidence that early tumor volume change on CT can predict pathological response and prognosis in non-small cell lung cancer patients receiving immunotherapy—an argument that volumetric assessment belongs at the center of response evaluation.

Turning volumetry from an aspiration into routine practice requires software that can delineate tumors without human tracing, and the engine of that capability is deep learning. The field’s foundational design is U-Net, introduced in 2015: an encoder–decoder architecture in which the encoder progressively compresses an image into abstract feature maps while the decoder expands them back into a voxel-by-voxel probability map marking tumor tissue. Its defining trick is the skip connection, which shuttles fine spatial detail from early encoder layers across the network bottleneck so that sharp boundaries survive compression. The three-dimensional extension, 3D U-Net, carried this logic to entire volumetric stacks, learning dense segmentation from sparse annotation—a natural fit for CT, where a tumor occupies a connected cluster of voxels across dozens of slices. Trained on expert-drawn contours with overlap-sensitive losses such as the Dice coefficient, these networks learn to separate malignant tissue from vessels, bronchi, pleura, and air. The clinical payoff is reproducibility: two radiologists handed the same scan will draw slightly different boundaries, while an algorithm returns the same contour every time.

From that foundation, the survey traces a decade of escalating sophistication. Attention mechanisms, popularized in medical imaging by Attention U-Net, force networks to suppress irrelevant anatomy and concentrate computational focus on suspicious regions, sharpening boundaries in the cluttered thoracic environment. UNet++ rebuilt the skip pathways as nested, densely connected branches that fuse features across multiple scales, better matching lesions ranging from sub-centimeter nodules to large invasive masses. Self-configuring frameworks such as nnU-Net automated the preprocessing and training choices that once demanded painstaking hand-tuning, and have become de facto baselines across segmentation benchmarks. Hybrid CNN–Transformer models such as TransUNet and Swin UNETR graft self-attention onto convolutional backbones: convolutions excel at local edges and textures, while attention layers relate distant patches of the image, supplying the long-range context needed for lobulated tumors sprawling across many slices. Multiscale, deeply supervised designs and recurrent residual variants were engineered specifically for lung tumors, and ensembles of 3D U-Nets have pushed nodule segmentation accuracy further. The survey organizes this crowded landscape by architecture, learning strategy, dataset, and intended clinical use, citing multicenter comparisons of Swin UNETR, nnU-Net, and TransUNet on lung tumor subtypes as evidence that architectural choice materially changes what a volume measurement is worth.

Anatomy alone, however, is not always decisive, and some of the most capable systems fuse CT with positron emission tomography. In PET/CT, the CT supplies structure while the PET signal reveals metabolism: malignant cells avidly consume a radioactive glucose analog, lighting up tumors whose borders are anatomically ambiguous. Multimodal networks exploit this pairing through spatial attention modules that learn, location by location, which modality to trust, recovering boundaries that either scan alone would blur. Beyond cleaner contours, the fusion yields metabolic volume biomarkers—measures such as metabolically active tumor volume that combine size with biological aggressiveness and have demonstrated value for risk stratification in non-small cell lung cancer. Publicly released whole-body FDG-PET/CT datasets with manually annotated tumor lesions, together with systematic benchmarks comparing architectures and training strategies on PET/CT volumes, are giving this multimodal branch of the field the standardized footing it long lacked.

None of this learning happens without data, and the survey pays sustained attention to the public resources anchoring the field. The Lung Image Database Consortium–Image Database Resource Initiative, known universally as LIDC-IDRI, remains the canonical repository for nodule detection, prized for its layered annotations from multiple radiologists. RIDER Lung CT occupies a different niche: it contains repeat scans of the same patients acquired minutes apart, the raw material for test–retest reproducibility studies that quantify how much apparent tumor change is biology versus measurement noise. NSCLC-Radiomics links CT-derived features to patient outcomes, enabling prognostic volumetric modeling. Yet each resource also exposes systemic weaknesses. Expert 3D annotation is slow and expensive, and radiologists genuinely disagree about tumor boundaries, especially where lesions abut the chest wall or blend into surrounding tissue; those disagreements propagate into training labels and, downstream, into the volumes being measured. Scanners, slice thicknesses, reconstruction kernels, and contrast protocols differ between hospitals, and models tuned to one acquisition style can falter on another—the domain-shift problem that shadows every medical AI deployment.

The survey’s central argument is that segmentation accuracy, however dazzling, is not the finish line; the question is whether the resulting volumes are stable enough to support clinical calls. Here, the evidence is accumulating. Multi-center, multi-observer reading studies of automated RECIST 1.1 and volumetric RECIST have found that machine-derived response assessments can rival the reliability of expert panels while removing much of the human variability. Volumetric CT growth-rate measurements have detected treatment effects in metastatic disease earlier than diameter-based criteria. In mesothelioma, where tumors wrap around the lung in shapes that defeat linear rules, fully automated deep learning volumetry has been validated against modified RECIST response criteria. In lung cancer specifically, early tumor volume change is being advanced as a novel CT indicator of pathological response and prognosis under immunotherapy, and volumetric measurements are being combined with radiomic texture features and blood biomarkers to anticipate pseudoprogression—the apparent growth that is actually immune-cell infiltration. Each of these advances rests on a segmentation mask a computer drew.

What separates these demonstrations from routine practice is a familiar list, which the authors dissect candidly. Annotated 3D datasets remain scarce relative to what deep learning craves, and annotation variability injects noise directly into the quantity clinicians hope to measure. Models are sensitive to scanner and acquisition protocol, so accuracy reported on one cohort rarely transfers unchanged to another. Interpretability is thin: a network offers no anatomical reasoning for the contour it draws, which complicates both clinician trust and regulatory approval. Multimodal integration—fusing CT with PET, clinical variables, and molecular data—remains technically immature. Most subtly, a high segmentation score does not guarantee volumetric reliability: a 5 percent apparent change in tumor volume between two scans may reflect boundary ambiguity rather than biology. Without test–retest validation and explicit uncertainty estimates, volumetric AI risks reproducing at scale the very inconsistency it was designed to eliminate.

The roadmap the authors chart is correspondingly forward-looking. Transformer-based volumetric models are expected to carry the architectural revolution that reshaped language processing into fully three-dimensional medical segmentation. Self-supervised and semi-supervised learning promise to exploit the enormous reservoir of unlabeled CT scans, teaching networks general lung anatomy before scarce expert contours are ever introduced—a direction embodied by recently reported lung CT foundation models trained on vast imaging collections for broad diagnostic use. Generative modeling could simulate tumor dynamics, synthesizing plausible growth and shrinkage sequences to augment scarce longitudinal data and train systems that predict response before it is visible. Perhaps most consequential for the clinic are uncertainty-aware systems that attach a confidence estimate to every contour, flagging cases where the algorithm is guessing so that human expertise is spent where it matters most. The authors also urge that volumetric biomarkers be validated against true clinical endpoints—pathological response, progression-free survival—rather than against segmentation scores, redefining success as decisions improved rather than pixels matched.

The stakes of getting this right are difficult to overstate. Global cancer statistics for 2022 place lung cancer among the most frequently diagnosed malignancies in the world and its deadliest, and low-dose CT screening programs are generating scan volumes that no human workforce can annotate the traditional way. If AI-driven volumetry matures, every follow-up scan could yield not a single diameter but a quantitative trajectory of tumor burden: earlier signals of response, earlier warnings of treatment failure, and measurements independent of which radiologist read the images. The Izmir survey is candid that the field stands closer to promise than to practice—the algorithms have largely learned to see, while medicine has not yet learned to trust. Closing that gap, the authors conclude, is less a computer-science problem than a clinical one, and the decisive experiments of the coming decade will unfold not in code repositories but in oncology wards, where a number that finally matches the true shape of a tumor could change what survival looks like.