Virtual Library
Start Your Search
Yohei Kawaguchi
Author of
-
+
P1.13 - Staging (ID 181)
- Event: WCLC 2019
- Type: Poster Viewing in the Exhibit Hall
- Track: Staging
- Presentations: 1
- Now Available
- Moderators:
- Coordinates: 9/08/2019, 09:45 - 18:00, Exhibit Hall
-
+
P1.13-09 - Utility of Artificial Intelligence in Estimation of the Histologic Type of Lung Cancer (Now Available) (ID 1038)
09:45 - 18:00 | Presenting Author(s): Yohei Kawaguchi
- Abstract
Background
With the development of artificial intelligence (AI), various activities using AI have become able to be performed without specialized knowledge. Computer-aided diagnosis systems using AI have also made great strides in decades. We examined the qualitative diagnostic ability to estimate the histologic type of lung cancer using an image recognition system that can be used online.
Method
We used 316 computed tomography (CT) images of lung cancer with solid component diameter less than 3 cm resected at our hospital. All images were trimmed at the tumor edge to increase the accuracy of machine learning by AI. Prepared images were classified by pathological diagnosis into adenocarcinoma (AD) group and non-adenocarcinoma (non-AD) group. 159 images were assigned to the training set and 157 were assigned to the test set. IBM watson studio; visual recognition app developed for image recognition was used for machine learning and judgment. The established algorithm by the training set was applied to the test set. The histologic type of which possibility calculated by AI was over 0.5 was defined as the AI answer.
Result
There were 93 AD and 66 non-AD in the training set and 92 AD and 65 non-AD in the test set. In the AD group, the median of the solid component diameter was 1.5 cm (0 - 2.9 cm). 21 images were pure ground-glass nodules, 122 images were part-solid ground-glass nodules and 42 images were consisted of solid component. In the non-AD group, the median of the solid component diameter was 2.0 cm (0.5 - 3.0 cm) and all images were consisted of solid component.
Of the 65 non-AD images in the test set, the AI answer was correct in all images (100%). However, of the 92 AD images in the test set, the AI answer was correct only in 49 images (53%). When the 47 AD images with dominant ground-glass opacity were analyzed, the AI answer was correct in 33 images (70%).
Conclusion
Although the CT image recognition using AI could accurately estimate the histologic type of lung cancer in tumors with dominant ground-glass opacity, it was difficult to distinguish solid AD from solid non-AD. The multimodal image analysis including enhanced CT and FDG-PET seems necessary.
-
+
P2.13 - Staging (ID 315)
- Event: WCLC 2019
- Type: Poster Viewing in the Exhibit Hall
- Track: Staging
- Presentations: 1
- Now Available
- Moderators:
- Coordinates: 9/09/2019, 10:15 - 18:15, Exhibit Hall
-
+
P2.13-14 - The Utility of Three-Dimensional CT for Prediction of Tumor Invasiveness in Clinical IA Lung Acenocarcinoma (Now Available) (ID 2478)
10:15 - 18:15 | Presenting Author(s): Yohei Kawaguchi
- Abstract
Background
In the evaluation of clinical T factor, there are some cases that the measurement of the solid component diameter of the tumor is difficult in conventional two-dimensional computed tomography (CT) because of heterogeneity and indistinctness of tumor density. So there is a problem in the preoperative estimation of the tumor invasiveness in these cases. Three-dimensional image analysis software, Synapse Vincent can quantify the volume of solid component of the lung nodule based on the CT value semi-automatically.The purpose of this study was to investigate the relationship between the histological grade of adenocarcinoma and the solid component volume by three-dimensional CT.
Method
We enrolled 195 cases of cIA adenocarcinoma resected at our hospital in 2017. Two observers measured the solid component diameter of the tumors after consultation.
The relationship of solid component diameter (2D), solid component volume (3D) and pathological subtypes (AIS, MIA or invasive cancer) were analyzed.
We additionally performed the same analysis with a focus on 57 cases (29.2%) in which we judged that 2D measurement of the tumor was difficult.
The cut-off value of each item was determined using the ROC curve.
Result
The number of AIS / MIA were 86 and of invasive cancer were 109 cases respectively. The median value of 3D was 442.2 mm3 (0-7044 mm3). About the prediction of invasive cancer by 2D, the sensitivity was 95.4% and the specificity was 64.0%. In the analysis of 3D, the sensitivity was 93.6% and the specificity was 69.6% assuming that the 3D cutoff value was 225 mm3. They were not statistically higher than that of 2D.
In subgroup analysis for 57 cases with difficulty in 2D measurement , when the cutoff value of 3D is 225mm3, the sensitivity is relatively good at 92.9% and the specificity 65.5%, and the accuracy is almost the same as the usual tumors with measurable solid components for invasive cancer prediction. In the analysis of 2D, the sensitivity was good at 92.9%, but the degree of specificity clearly decreased at 44.8%, and the diameter of the solid component tended to be overestimated.
Conclusion
Measurement of the solid component diameter by two-dimensional CT tends to over-estimate shadows that are difficult to measure. Three-dimensional CT, semi-automatic measurement of solid component volume, can be performed easily, and the usefulness of it was suggested, especially in cases with the tumors which are difficult to measure the solid component diameter by two-dimensional CT.
