Performance evaluation tools for zone segmentation and classification

Current Medical Imaging Reviews 14 1 — Multimed Tools Appl:1— Comp Methods Prog Biomed 1 — Magn Reson Imaging 31 7 — Pearson Educ — Med Phys 7. IEEE Access — IEEE Access. InInternational conference on neural information processing Dec 13 pp. Springer, Cham. Electron Lett 44 13 — Procedia Computer Science — BJU Int 3 — Magn Reson Med 77 3 — Deep learning nature — Lehmann EL, Casella G Theory of point estimation.

Comput Biol Med — InTwelfth international conference on quality control by artificial vision Apr 30 Vol. Eur J Radiol — Lim JS Two-dimensional signal and image processing. Prentice hall, Englewood Cli s, NJ, pp. Liu Y, An X A classification model for the prostate cancer based on deep learning. InMedical imaging computer-aided diagnosis Feb 26 Vol. Biomedical Signal Processing and Control. Magn Reson Med 36 3 — Med Image Anal 22 1 — Med Phys 37 4 — Mohammadi M, Nabavi S n. Improvement in auto-contouring approaches using region growing segmentation for prostate cancer radiotherapy.

Phys Med Biol 57 12 — Perona P, Malik J Scale-space and edge detection using anisotropic diffusion. Magn Reson Imaging 30 10 — Br J Radiol 71 — InHandbook of biomedical image analysis pp. Springer, Boston, MA. Sarkar S, Das S A review of imaging methods for prostate cancer detection: supplementary issue: image and video acquisition and processing for clinical applications.

Save to Library Save. Create Alert Alert. Share This Paper. Methods Citations. Figures and Tables from this paper. Citation Type.

Has PDF. It extends their approach by providing scheme in which result zones belonging to the same algorithm-specific set of metrics for overlap analysis, ground-truth zone can be merged using an optional RLE and polygonal representation of regions and in- merge option. While it remains a zone-based evalua- troduces type-matching to evaluate zone classification. This merge-option is config- This is followed by the implementation details of zone urable and evaluation tool can be used with both merge matching, the merge capability to analyze substitution lenient and no-merge stricter options.

PETS Overview tion. It works like zone segmentation with both merge and no-merge options, but requires an additional con- The evaluation of various detection, segmentation straint of zone-type zone classification to be matched and zone-classification algorithms like page segmen- before a merge or an overlap can be established.

However, each of these algorithms require a category of zone detection of particular types of con- different evaluation strategy for detailed analysis. Ground-truthing all regions in a document is not expected and any result zones overlapping with pix- 2. Evaluation of zone classification algorithms require One challenge that should be dealt with is when consistent zone types assuming a pre-established cor- there is a difference in scope between the ground truth respondence.

This is the simplest evaluation scheme and the results. For example, a dataset may only be where result zone is penalized only when its recog- annotated with the content the user is interested in nized type is not equivalent to the ground-truth zone evaluating.

A collection may be annotated with con- type and does not consider location. This is fine for detection, where we expect only the regions of interest are evaluated. This helps in identified regions to be of a certain type, but is not nec- evaluating a subset of region-types when the datasets essarily appropriate for segmentation. A segmentation are significantly unbalanced [1].

To address this problem, PETS guable step in a document processing pipeline. While it is evi- dent that two different style zones should always be 3. Implementation returned as distinct zones, splitting a text-zone along the direction of text, for example at paragraph or line 3.

In order to avoid this con- PETS requires three sets of files as inputs: image, fusion, accuracy of page segmentation algorithms has ground-truth and result files.

The ground-truth and re- been calculated as the percentage of ground-truth text- sult files follow the GEDI XML format specification [8]. The drawback of ument interface for scanned text documents [14]. All segmentation and clas- many-to-one and many-to-many. This XML file can also be visualized in GEDI along with the corre- sponding image file to analyze the segmentation Figure 2: a One-to-one b One-to-many c Many-to- or classification algorithm under study.

The overall result file contains the matching scores Comparing the cardinal of each union, following ob- of all zones, confusion matrices and a summarized servations are made: result with precision, recall and F1 scores. Missed: Z g having no corresponding Zr 3. False Alarm: Zr having no corresponding Z g 3. Simple Match: Cardinality of each of Z g and Zr is Zones are represented either as rectangular with one.

In case of overlapping zones, pixels can be associated with a 4. Multi-match: When at least one of Z g or Zr is zone using run-length encoding. The proposed PETS greater than 1 and none are zero. Since pixel-based overlap computation PETS evaluates various algorithms depending on is an expensive operation, especially when zones are the usage options described below: polygonal, a bounding box overlap is computed first. Assuming zone cor- based overlap is computed.

Regions which have a set of MST values above user- Figure 3 shows the evaluation result of a docu- specified thresholds are said to have overlapped, oth- ment image on our voronoi based segmentation erwise the values are reset to zero.

Therefore, manual segmentation of brain tumors from magnetic resonance MR images is a challenging and time-consuming task. In addition, an automated brain tumor classification from an MRI scan is non-invasive so that it avoids biopsy and make the diagnosis process safer.