Automatic Detection of Mandibular Fractures in Panoramic Radiographs Using Deep Learning

Abstract

Mandibular fracture is one of the most frequent injuries in oral and maxillo-facial surgery. Radiologists diagnose mandibular fractures using panoramic radiography and cone-beam computed tomography (CBCT). Panoramic radiography is a conventional imaging modality, which is less complicated than CBCT. This paper proposes the diagnosis method of mandibular fractures in a panoramic radiograph based on a deep learning system without the intervention of radiologists. The deep learning system used has a one-stage detection called you only look once (YOLO). To improve detection accuracy, panoramic radiographs as input images are augmented using gamma modulation, multi-bounding boxes, single-scale luminance adaptation transform, and multi-scale luminance adaptation transform methods. Our results showed better detection performance than the conventional method using YOLO-based deep learning. Hence, it will be helpful for radiologists to double-check the diagnosis of mandibular fractures.

Keywords: YOLO; YOLO v4; deep learning; image processing; mandibular fracture; multi-scale luminance adaptation transform (MLAT); object detection; panoramic radiography; single-scale luminance adaptation transform (SLAT).

Figure 1

Examples of mandibular fracture: (a) anatomic area of mandibular fracture [2], (b) mandibular fractures (red boxes) in panoramic radiographs, (c) linear fracture on parasymphysis area, and (d) shear fracture on angle area.

Figure 2

The result of breast mass detection by Mohammed et al. [10]: (a) Ground truth of mass (red circle), (b) detection by Mohammed et al. (pink box), (c) ground truth of a malignant case (red circle), and (d) detection by Mohammed et al. (green box). Reprinted with permission from ref. [10]. Copyright 2018, Elsevier.

Figure 3

The result of cyst detection (green boxes) by Yang et al. [12]: (a) Odontogenic keratocyst and (b) dentigerous cyst, both detected by Yang et al.

Figure 4

YOLO prediction feature map of the mandibular fracture panoramic radiograph: Red box is a feature map tensor of the mandibular fracture.

Figure 5

The brief structure of YOLO v4.

Figure 6

YOLO v4 backbone structure: Cross-stage-partial-connection-Darknet53.

Figure 7

YOLO v4 neck structure: Spatial pyramid pooling layer.

Figure 8

YOLO v4 path aggregation network neck and head structure.

Figure 9

Single-scale luminance adaptation transform.

Figure 10

The comparison between original panoramic radiograph, single-scale luminance adaptation transform (SLAT), and multi-scale luminance adaptation transform (MLAT) panoramic radiograph: (a) Original, (b) SLAT, and (c) MLAT.

Figure 11

The comparison between different four modules for mandibular fracture detection (white boxes): (a) Diagnosed radiographs (blue lines) by a radiologist, (b) Original module, (c) Gamma modulation module, (d) Luminance adaptation transform module, and (e) The proposed module.

Figure 12

The proposed method’s block diagram of the diagnosis process in YOLO v4.

Figure 13

The two-class and six-class comparison of correct diagnosed distribution (red circles) of mandibular fractures: (a) two-class single-scale luminance adaptation transform (SLAT), (b) six-class SLAT, (c) two-class multi-scale luminance adaptation transform (MLAT), and (d) six-class MLAT.

Figure 14

The two-class and six-class comparison of undiagnosed distribution (red triangles) of mandibular fractures: (a) two-class single-scale luminance adaptation transform (SLAT), (b)six-class SLAT, (c) two-class multi-scale luminance adaptation transform (MLAT), and (d) six-class MLAT.

Figure 14

The two-class and six-class comparison of undiagnosed distribution (red triangles) of mandibular fractures: (a) two-class single-scale luminance adaptation transform (SLAT), (b)six-class SLAT, (c) two-class multi-scale luminance adaptation transform (MLAT), and (d) six-class MLAT.

Figure 15

The comparison of multi-scale luminance adaptation transform (MLAT), single-scale luminance adaptation transform (SLAT), and MLAT and SLAT in two-class.

Figure 16

The comparison of multi-scale luminance adaptation transform (MLAT), single-scale luminance adaptation transform (SLAT), and MLAT and SLAT in six-class.

Figure 17

The comparison of two-class and six-class multi-scale luminance adaptation transform and single-scale luminance adaptation transform modules.

Figure 18

The mandibular fractures diagnosis comparison: (a) diagnoses mandibular fractures (orange lines) by radiologist, (b) mandibular fracture detection (red boxes) of two-class multi-scale luminance adaptation transform (MLAT) and single-scale luminance adaptation transform (SLAT) module, and (c) mandibular fracture detection (red boxes) of six-class MLAT and SLAT module.

Figure 19

The mandibular fractures diagnosis comparison: (a) diagnoses mandibular fractures (orange lines) by radiologist, (b) mandibular fracture detection (red boxes) of two-class multi-scale luminance adaptation transform (MLAT) and single-scale luminance adaptation transform (SLAT) module, and (c) mandibular fracture detection (red boxes) of six-class MLAT and SLAT module.

Figure 20

The mandibular fractures diagnosis comparison: (a) diagnoses mandibular fractures (orange lines) by radiologist, (b) mandibular fracture detection (red boxes) of two-class multi-scale luminance adaptation transform (MLAT) and single-scale luminance adaptation transform (SLAT) module, and (c) mandibular fracture detection (red boxes) of six-class MLAT and SLAT module.