Caries is a gradient, not a boundary: Detection rather than segmentation is the appropriate deep learning approach

Artificial intelligence research in dental radiology continues to grow rapidly, and automated detection of dental caries has become a major focus.1,2 Although deep learning can be applied in many different ways, recent caries AI research has focused mainly on two approaches: detection and segmentation.3,4 Segmentation appears appealing because it produces a lesion contour on the radiograph and creates an impression of precision. However, segmentation assumes that dental caries has a definable boundary. This assumption is biologically incorrect. Caries is not an anatomical object; it is a continuous process of mineral loss. Radiographically, the demineralization associated with caries often manifests as a gradual change in density rather than a sharply defined lesion boundary.5 When segmentation is applied, the model is forced to draw a boundary that does not exist.

Radiographically, early caries begins as subsurface demineralization. As mineral content decreases, the demineralized zone becomes radiolucent, and toward the periphery the radiolucency gradually transitions into sound enamel or dentin.5 This transition never occurs at a single pixel in bite-wing, periapical, or panoramic images; the lesion fades across multiple pixels and blends gradually into the surrounding tooth structure. An “exact” contour of caries therefore does not reflect clinical reality (Fig. 1A).

Fig. 1. A. Bitewing and panoramic images show a proximal caries on the distal surface of the right mandibular first molar. The lesion fades across multiple pixels and blends into the surrounding tooth structure. B. Segmentation of dental caries using a closed curve. The boundary is unclear and drawn arbitrarily C. Detection of dental caries using a square box. The box includes the entire area of caries.

Segmentation, however, requires a pixel-level binary decision: each pixel must be labeled as caries or non-caries. This binary framework conflicts with the biological nature of demineralization, which progresses along a continuous spectrum rather than at a discrete boundary. A segmentation mask may appear visually precise, but the precision is optical rather than biological. By forcing a sharp edge onto a process without sharp limits, segmentation creates a categorical output that conveys a false sense of accuracy (Fig. 1B).

Segmentation also lacks reproducibility. To construct segmentation datasets, annotators manually trace lesion borders. Because the border is not visible, annotators must estimate where to draw the line. Two experienced clinicians may create different boundaries for the same lesion—not because one is right and the other wrong, but because a true boundary cannot be observed. Segmentation does not remove subjectivity; it visualizes it. The uncertainty becomes part of the dataset before training even begins, and the model learns annotation subjectivity rather than the biology of disease.

Object detection aligns more closely with clinical reasoning. In routine practice, clinicians do not delineate the exact margins of a carious lesion on a radiograph. The essential task is to recognize whether caries is present and, if so, estimate the general extent of involvement to inform treatment planning rather than to define a precise boundary. Detection models highlight regions of concern with bounding boxes and provide a clinically useful indication of lesion presence and approximate size, without implying that the model can determine the true biological extent of caries.6 The model guides attention while preserving the clinician’s judgment.

Detection is also more practical in dataset generation. Bounding boxes can be annotated quickly and consistently, enabling the creation of larger and more diverse datasets. Larger datasets improve generalizability and model stability across patient groups, radiograph types, and imaging systems. Segmentation requires pixel-level annotation, which is labor-intensive and heavily influenced by the annotator’s interpretation. This leads to smaller datasets and reduces clinical applicability (Fig. 1C).

From a regulatory standpoint, detection is safer. Bounding boxes indicate a region of interest, whereas segmentation masks visually imply lesion extent.7 A model that appears to define disease boundaries may be interpreted as providing diagnostic or treatment guidance, increasing medical-legal risk. Regulatory agencies emphasize safety, interpretability, and clarity of responsibility. Detection aligns naturally with these goals; segmentation may give the impression of determinism that cannot be supported.

Segmentation remains common simply because it aligns with how engineering models are traditionally built: they are designed to draw boundaries. Clinicians, however, do not diagnose caries by locating an exact edge, because such an edge does not exist radiographically or biologically. This fundamental difference in perspective-engineers seeking a boundary to outline and clinicians recognizing that no such boundary is present-helps explain why segmentation often fails to translate into meaningful clinical value.

Caries diagnosis is pattern recognition.8 Clinicians identify gradual changes in density and shape and consider contextual factors, including patient history and adjacent anatomy. When residents learn to read radiographs, they are not asked to draw lesion borders. They are taught to recognize patterns, interpret uncertainty, and make decisions. Detection reflects this approach by identifying suspicious areas while allowing clinicians to interpret significance.

If grayscale intensity is sampled across a carious lesion, the plot shows a gradual slope representing progressive demineralization. There is no abrupt drop corresponding to an anatomical boundary. Segmentation forces this slope into a vertical cliff. The cliff represents the constraints of the method, not the biology of disease.

Artificial intelligence should support clinical decision-making, not redefine morphology. For a disease that exists as a gradient, a model that forces a binary boundary is misaligned with reality. Dental caries is a gradient, not a boundary. For this reason, detection rather than segmentation is the clinically appropriate approach for AI-based caries diagnosis.