Analysis of Bone Scans in Various Tumor Entities Using a Deep-Learning-Based Artificial Neural Network Algorithm-Evaluation of Diagnostic Performance.

Wuestemann, Jan; Hupfeld, Sebastian; Kupitz, Dennis; Genseke, Philipp; Schenke, Simone; Pech, Maciej; Kreissl, Michael C.; Grosser, Oliver S.

doi:10.3390/cancers12092654

by Jan Wuestemann, Sebastian Hupfeld, Dennis Kupitz, Philipp Genseke, Simone Schenke, Maciej Pech, Michael C. Kreissl, Oliver S. Grosser

Abstract:

The bone scan index (BSI), initially introduced for metastatic prostate cancer, quantifies the osseous tumor load from planar bone scans. Following the basic idea of radiomics, this method incorporates specific deep-learning techniques (artificial neural network) in its development to provide automatic calculation, feature extraction, and diagnostic support. As its performance in tumor entities, not including prostate cancer, remains unclear, our aim was to obtain more data about this aspect. The results of BSI evaluation of bone scans from 951 consecutive patients with different tumors were retrospectively compared to clinical reports (bone metastases, yes/no). Statistical analysis included entity-specific receiver operating characteristics to determine optimized BSI cut-off values. In addition to prostate cancer (cut-off = 0.27\%, sensitivity (SN) = 87\%, specificity (SP) = 99\%), the algorithm used provided comparable results for breast cancer (cut-off 0.18\%, SN = 83\%, SP = 87\%) and colorectal cancer (cut-off = 0.10\%, SN = 100\%, SP = 90\%). Worse performance was observed for lung cancer (cut-off = 0.06\%, SN = 63\%, SP = 70\%) and renal cell carcinoma (cut-off = 0.30\%, SN = 75\%, SP = 84\%). The algorithm did not perform satisfactorily in melanoma (SN = 60\%). For most entities, a high negative predictive value (NPV ≥ 87.5\%, melanoma 80\%) was determined, whereas positive predictive value (PPV) was clinically not applicable. Automatically determined BSI showed good sensitivity and specificity in prostate cancer and various other entities. Particularly, the high NPV encourages applying BSI as a tool for computer-aided diagnostic in various tumor entities.

Reference:

Analysis of Bone Scans in Various Tumor Entities Using a Deep-Learning-Based Artificial Neural Network Algorithm-Evaluation of Diagnostic Performance. (Jan Wuestemann, Sebastian Hupfeld, Dennis Kupitz, Philipp Genseke, Simone Schenke, Maciej Pech, Michael C. Kreissl, Oliver S. Grosser), In Cancers, volume 12, 2020.

Bibtex Entry:

@article{wuestemann_analysis_2020,
	title = {Analysis of {Bone} {Scans} in {Various} {Tumor} {Entities} {Using} a {Deep}-{Learning}-{Based}  {Artificial} {Neural} {Network} {Algorithm}-{Evaluation} of {Diagnostic} {Performance}.},
	volume = {12},
	issn = {2072-6694 2072-6694 2072-6694},
	doi = {10.3390/cancers12092654},
	abstract = {The bone scan index (BSI), initially introduced for metastatic prostate cancer,  quantifies the osseous tumor load from planar bone scans. Following the basic idea  of radiomics, this method incorporates specific deep-learning techniques (artificial  neural network) in its development to provide automatic calculation, feature  extraction, and diagnostic support. As its performance in tumor entities, not  including prostate cancer, remains unclear, our aim was to obtain more data about  this aspect. The results of BSI evaluation of bone scans from 951 consecutive  patients with different tumors were retrospectively compared to clinical reports  (bone metastases, yes/no). Statistical analysis included entity-specific receiver  operating characteristics to determine optimized BSI cut-off values. In addition to  prostate cancer (cut-off = 0.27\%, sensitivity (SN) = 87\%, specificity (SP) = 99\%),  the algorithm used provided comparable results for breast cancer (cut-off 0.18\%, SN  = 83\%, SP = 87\%) and colorectal cancer (cut-off = 0.10\%, SN = 100\%, SP = 90\%). Worse  performance was observed for lung cancer (cut-off = 0.06\%, SN = 63\%, SP = 70\%) and  renal cell carcinoma (cut-off = 0.30\%, SN = 75\%, SP = 84\%). The algorithm did not  perform satisfactorily in melanoma (SN = 60\%). For most entities, a high negative  predictive value (NPV ≥ 87.5\%, melanoma 80\%) was determined, whereas positive  predictive value (PPV) was clinically not applicable. Automatically determined BSI  showed good sensitivity and specificity in prostate cancer and various other  entities. Particularly, the high NPV encourages applying BSI as a tool for  computer-aided diagnostic in various tumor entities.},
	language = {eng},
	number = {9},
	journal = {Cancers},
	author = {Wuestemann, Jan and Hupfeld, Sebastian and Kupitz, Dennis and Genseke, Philipp and Schenke, Simone and Pech, Maciej and Kreissl, Michael C. and Grosser, Oliver S.},
	month = sep,
	year = {2020},
	pmid = {32957650},
	pmcid = {PMC7565494},
	keywords = {bone metastases, bone scan, bone scan index, deep learning, radiomics}
}