UNet and MobileNet CNN-based model observers for CT protocol optimization: comparative performance evaluation by means of phantom CT images

Federico Valeri; Maurizio Bartolucci; Elena Cantoni; Roberto Carpi; Evaristo Cisbani; Ilaria Cupparo; Sandra Doria; Cesare Gori; Mauro Grigioni; Lorenzo Lasagni; Alessandro Marconi; Lorenzo Nicola Mazzoni; Vittorio Miele; Silvia Pradella; Guido Risaliti; Valentina Sanguineti; Diego Sona; Letizia Vannucchi; Adriana Taddeucci

doi:10.1117/1.JMI.10.S1.S11904

7 March 2023 UNet and MobileNet CNN-based model observers for CT protocol optimization: comparative performance evaluation by means of phantom CT images

Federico Valeri, Maurizio Bartolucci, Elena Cantoni, Roberto Carpi, Evaristo Cisbani, Ilaria Cupparo, Sandra Doria, Cesare Gori, Mauro Grigioni, Lorenzo Lasagni, Alessandro Marconi, Lorenzo Nicola Mazzoni, Vittorio Miele, Silvia Pradella, Guido Risaliti, Valentina Sanguineti, Diego Sona, Letizia Vannucchi, Adriana Taddeucci

Author Affiliations +

Journal of Medical Imaging, Vol. 10, Issue S1, S11904 (March 2023). https://doi.org/10.1117/1.JMI.10.S1.S11904

Fig. 1

View Large | Download | View In Article Context

Lateral view (a) and top view (b) of one of the two blocks containing five inserts filled with iodinated contrast media.

Table 1

Selected subensembles from the original image dataset characterized by insert (object) size and contrast.

No. images	object diameter	contrast
	d (mm)	C (HU)
10,000	Homogeneous images	—
3000	3	45
3000	3	55
3000	4	45
3000	4	55
2800	5	45
2800	5	55
1200	6	45
1200	7	45

Fig. 2

View Large | Download | View In Article Context

Example of reconstructed images (iterative reconstruction technique) with (a) 4 mm and (b) 7 mm inserts at the lower contrast (45 HU); it is noticeable that object visibility depends both on size and CTDIvol.

Fig. 3

View Large | Download | View In Article Context

Screenshot of the software interface developed to collect the human observer response to CT images visual inspection.

Fig. 4

View Large | Download | View In Article Context

Schematic illustration of the developed UNet-based CNN architecture.

Fig. 5

View Large | Download | View In Article Context

Schematic illustration of the MobileNetV2-based CNN architecture used for the classification task.

Fig. 6

View Large | Download | View In Article Context

Schematic illustration of the MobileNetV2-based CNN architecture developed for the localization task.

Table 2

Selected thresholds for the LROC curves computation for different insert diameters.

Insert diameter (mm)	3	4	5	6	7
Threshold (mm)	2.3	2.3	2.5	3.0	3.5

Fig. 7

View Large | Download | View In Article Context

Human observer performances quantified by LAUC versus CTDIvol, for different object sizes, contrasts (left: 45 HU, right: 55 HU), and reconstruction techniques (a) FBP and (b) IR.

Fig. 8

View Large | Download | View In Article Context

Overall MO and human observer performances quantified by LAUC versus CTDIvol, for the two reconstruction techniques (a) FBP and (b) IR, with associated standard errors.

Fig. 9

View Large | Download | View In Article Context

Comparison of human observer and MO LAUCs versus CTDIvol for the images with an object size of 4 mm, C=45 HU, and IR reconstruction, with associated standard errors.

Table 3

MAPE between human observer and MO LAUCs for full IR and FBP datasets and a representative case (Fig. 9).

CNN	FBP	IR	4 mm, IR, C=45 HU
UNet	2.36	1.43	1.19
MobileNetV2	2.49	1.52	2.48

Fig. 10

View Large | Download | View In Article Context

MOs localization accuracy metric versus each of the independent parameters, from left to right: insert diameter, insert contrast, CTDIvol, and reconstruction techniques.