Improvement in matching lesions in dual-view mammograms using a geometric model

Abstract

Objectives: To evaluate the effectiveness of a geometric model (GM) as an adjunctive tool for radiologists to match lesions between craniocaudal (CC) and mediolateral (MLO) views.

Methods: A retrospective study was conducted on 711 patients who underwent mammography from January 2016 to August 2018. Two senior radiologists used bounding boxes to delineate lesions as the reference standard, calculated the absolute error (the shortest distance from the lesion center to the predicted curve) of GM, and compared it with the annular band (AB) and straight strip (SS) methods. Four radiologists of varying seniority levels were tasked with localizing the corresponding lesion in MLO view using a bounding box, based on the given lesion in CC views, and recording reading time per case with or without GM assistance. The Dice coefficient was used to evaluate the overlap between the bounding box and the reference standard.

Results: Overall, 499 calcification and 212 mass pairs were evaluated. GM outperformed both AB and SS, yielding a median absolute error of 3.03 mm (IQR 1.45-5.55 mm) versus 5.78 mm (IQR 2.44-10.71 mm) for AB and 4.59 mm (IQR 1.91-8.19 mm) for SS (P < 0.001). With GM assistance, all four radiologists achieved improved Dice coefficients and reduced reading times (all P < 0.001). Stratified analysis by lesion conspicuity demonstrated that GM assistance significantly enhanced Dice coefficients for all radiologists in the low-conspicuity group and improved matching consistency for junior radiologists.

Conclusion: The geometric model holds substantial promise as a valuable tool to assist radiologists in more effectively localizing lesions in ipsilateral mammograms, thereby potentially enhancing diagnostic accuracy and efficiency.

Keywords: Breast; Craniocaudal; Image matching; Mammography; Mediolateral oblique.

Fig. 1

Flow diagram showing patient selection details

Fig. 2

The CC views versus the MLO views of the radial distances (A) and the axial distances (B). Strong correlations between CC and MLO views were observed for both radial (r = 0.917) and axial (r = 0.923) distances

Fig. 3

Box plots depict the distribution of absolute errors (AE) across subgroups (N = 711 lesions), highlighting dispersion and skewness patterns. (A) AE for geometric model (GM), annular band (AB), and straight strip (SS) methods. (B) AE differences between high-density and low-density breasts. (C) AE differences between outer and inner breast quadrants. (D) AE variation across upper and lower breast quadrants

Fig. 4

The BI-RADS characteristics of the masses

Fig. 5

Reader performance enhancement with geometric model (GM) assistance. Dice coefficient ranges from 0 to 1, with 0 denoting no overlap and 1 denoting complete overlap. Box plots show that Dice coefficients improved for all readers (P < 0.05)

Fig. 6

Representative lesion localization performance with/without GM assistance. (A-B) FFDM images showing a target mass (red bounding box) in CC (A) and MLO (B) views. (C-D) Corresponding digital breast tomosynthesis (DBT) slices confirming lesion spatial consistency. (E-G) Manual MLO-view localizations by the four radiologists without GM assistance: R3 (Dice = 0.91, correctly identified, E); R1 (false correspondence, F); R2 and R4 (false correspondence, G). (H) A predicted curve provided by GM (green curve) passes through the mass with an AE of 3.12 mm. All radiologists achieved accurate localization with GM: Dice coefficients 0.86 (R1), 0.88 (R2), 0.92 (R3), 0.87 (R4)